This week, compiled:

• The Big Story: Anthropic replaced git in its monorepo with an in-house system prepared for agents, Scion, after merges grew from about 2,500 a day in May to over 30,000 in September

• The Release: OpenCode 2.0 runs the agent as one shared background service, turns about 80 built-in features into plugins, and stops 1.x plugins from working

• The Trend: Meta, Reflection and DeepSeek released open weights on October 15, 16 and 17, under three different licenses: the Muse Model License, Apache 2.0 and MIT

If you only read one thing this week: Anthropic says Claude now writes about 90% of Claude Code’s own code, and the team published how it reviews that code. Between June and September, merged PRs (pull requests) to Claude Code more than doubled, and the share that got reverted fell by more than 40%. The share of PRs that needed five or more review rounds went from 8.9% to 1.5%. The idea I would take to my own teams is the risk tier on human approval. A low-risk change needs no human approval once the checks are green, a medium-risk change needs one person and a high-risk change needs two. Reviews also stop after three rounds, and a re-review looks only at the new diff. Every number here is company-measured and the post is on an Anthropic-hosted blog, so I read it as a process description, not a benchmark. The Postgres 19 delay further down shows the same problem from the other side. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• Anthropic replaced git in its monorepo after merges per day grew about twelvefold in four months: in a post on its engineering blog (October 13), Anthropic says merges went from about 2,500 a day in May to over 30,000 in September. The main-branch history is nearly 100 GB, and CI (continuous integration) makes about 25 million checkouts a day. The in-house replacement, Scion, retired several hundred git mirrors. Anthropic reports about 200x less CPU and memory than the mirror setup, and checkouts that take seconds instead of minutes. Seven people built Scion and its review tool, and the migration took five months. All of these numbers are the company’s own. A former Meta engineer argues that Scion’s derived-data design closely follows Meta’s Mononoke without credit, which is one person’s claim. Read more →

• OpenCode 2.0 runs the agent as one shared background service: on October 12 Anomaly published the OpenCode 2.0 announcement. Last week I could not find it, and all I had was the upgrade command in v1.19.0. The terminal, web and desktop apps now connect to one background service, and about 80 built-in features became internal plugins with hot reload. Code Mode turns MCP servers and OpenAPI endpoints into functions that the agent calls from one script, and every MCP server is opted in by default. Subagents and shell commands can run in the background, and a subagent can use a different model. The upgrade breaks things: 1.x plugins stop working and the server API is completely different. It happens only when you run opencode upgrade --major. On the 1.x line, v1.19.1 (October 15) removed session sharing, and v1.19.2 keeps 1.x installs on 1.x, although Homebrew Core now tracks 2.x. Read more →

• Meta, Reflection and DeepSeek released open weights on three consecutive days, under three different licenses: on October 15 Meta released Muse Spark Open, an 816B-parameter MoE (mixture of experts) with 42B active parameters and a 1M-token context, under the Muse Model License with limited commercial terms. Meta’s post gives no coding benchmark scores. On October 16 Reflection released the Beam weights under Apache 2.0, the release that Edition 52 could only report as planned. On October 17 DeepSeek shipped V4.1-Flash-1017 under MIT. It is a minor update, and DeepSeek’s own model card puts it at 42.3 on Terminal-Bench 4.0, up from 31.2, with the API price unchanged. DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later with it. By October 17 Cline’s model catalog already made Muse Spark Open and Beam the default model for some providers. Read more →

• Gemini 4 Argon is generally available, and its API price doubles on December 1: on October 13 Google made Argon generally available in the Gemini API and the Gemini app for Pro and Ultra subscribers, and began rolling it out in Antigravity. In Edition 51 only cyber defenders had access. The same day GitHub added it to Copilot for Pro+, Max, Business and Enterprise. Google’s pricing page lists $2 / $10 per million input / output tokens through November 30, then $4 / $20 from December 1, so the introductory price I reported in Edition 51 now has an end date. Argon is not on the API’s free tier. Google’s own benchmark table puts Argon behind Claude Opus 5.5 on Terminal-Bench 4.0, at 57.4% against 66.4%. Read more →

• Neuralwatt’s data shows that the harness decides how often the prompt cache breaks: Neuralwatt analyzed coding-agent traffic on its own inference service, and as an inference provider it has a commercial interest. Uncached input ranged from 1.6% for deepseek-harness to 11.1% for Node.js apps. The cache broke about once in 100 turns with deepseek-harness, and about three times in 100 with OpenCode and Pi. Requests with under 10% cached tokens used about 91% more energy on DeepSeek V4.1 Flash. In a lab test, OpenCode 1.x broke the cache with a plan-mode reminder, and 2.x fixed it. Neuralwatt estimates that past 128K tokens, one turn that re-reads over half its prompt can cost as much prefill as about 600 normal turns. Read more →

• Uber gives each engineer a monthly AI spend allocation and routes its models through one gateway: according to The Pragmatic Engineer’s look inside Uber’s AI infrastructure (October 14), a manager can approve extra spend for the month, and a plugin called CostCheck suggests ways to lower cost across sessions. Since 2023 Uber has kept its AI tooling independent of any one model or harness. A Model Gateway switches between Anthropic, OpenAI, other vendors and open models that Uber runs itself, and it tracks spend per user. A layer called AIFX lets engineers move between Claude Code, Codex and OpenCode and keep their settings. The piece does not say how Uber picks a model for each task. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for moving an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

GitHub makes agentic autofix and AI Scan for pull requests generally available

GitHub — October 15, 2026

Both features reached general availability on October 15, and both are billed in Code Security AI Credits. With agentic autofix you assign a code scanning alert to Copilot, or up to 25 alerts at once. Copilot explores the codebase, writes a fix, re-runs CodeQL to check it and opens a draft PR. It needs GitHub Code Security or Advanced Security on Enterprise Cloud, and it also uses Actions minutes. AI Scan adds LLM findings on pull requests for languages that CodeQL covers poorly, such as PHP, Shell, Terraform and Dockerfiles. Its findings are advisory and never block a merge. It is free for public repositories and needs no Copilot license.

Claude Code v2.1.299 to v2.1.302 compact idle sessions before the prompt cache expires

Anthropic — October 14, 15 and 16, 2026

The useful change is idle compaction in v2.1.299. When a conversation over 200K tokens sits idle for about 54 minutes, Claude Code compacts it before its one-hour prompt cache expires, so the next message does not reprocess the whole conversation. You can turn it off with "idleCompaction": false. v2.1.300 splits cache writes into 5-minute and 1-hour token counts in its OpenTelemetry events. It also gzip-compresses large requests to api.anthropic.com, even with telemetry off. v2.1.301 adds a setting that caps how long an unattended session waits at a usage limit. v2.1.302 fixes a prompt-cache miss when tools load mid-session, and it announces that standalone Claude Design closes on December 14, 2026. Neuralwatt’s numbers above show why the cache is worth protecting.

Codex CLI 0.163.0 lets you steer a running answer, and GPT-5.5 leaves Codex for ChatGPT sign-ins

OpenAI — October 14 and 15, 2026

In 0.163.0 (October 15) you can send a follow-up while Codex is still responding, and it changes course. /config is now searchable and applies changes live, the /keymap leader key is configurable, and you can restart a managed background server from /daemon. Read-only tools such as skills, memory and history now run in parallel. The network proxy checks permission before the DNS lookup and blocks credential aliases that got around inspection hooks. Separately, OpenAI’s October 14 changelog retires GPT-5.5 from Codex for users who sign in with ChatGPT. API-key users are not affected. Admins need to update defaults, scheduled tasks and scripts that pin gpt-5.5, and OpenAI points them to gpt-6.1-sol. Edition 52 covered 0.160.1 to 0.162.1.

Antigravity CLI 1.3.4 to 1.4.2 adds an Auto mode with a second agent as reviewer

Google — October 13 to 18, 2026

Google shipped six releases. In 1.3.5 (October 14), Auto mode approves edits and safe commands and asks before irreversible actions. Any command that would leave the terminal sandbox goes to a separate agent for review first. It works on Linux and macOS, and Windows comes later. 1.3.6 makes Claude Opus 5.5 and Sonnet 5.5 available through Application Default Credentials. It also lets the Auto-mode reviewer block uploads of sensitive data, and fixes a hang when the model connection stalls. 1.4.0 to 1.4.2 add terminal status reporting, rewrite agy plugin validate and fix smaller issues. Edition 52 covered 1.2.17 to 1.3.3.

GitHub Copilot: HydraFusion Max, AI Credits in /usage and three models retiring on November 16

GitHub — October 12 to 15, 2026

HydraFusion, the multi-model orchestrator I covered as a research preview in Edition 51, now has a Max option. It runs the same orchestration at higher reasoning effort where depth matters, with Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna in the pool, and it now works in Copilot CLI too. The changelog gives no price and no savings figure. Copilot CLI v1.0.97 shows billed AI Credits in /usage for token-billed sessions, and v1.0.96 shows in the timeline who made each permission decision. GitHub also set a retirement date: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra leave all Copilot experiences on November 16, 2026.

Kiro adds a local command sandbox, and Kimi K3 at a 2.4x credit multiplier

Kiro — October 12 to 15, 2026

IDE 1.3.0 (October 13) adds a local command sandbox that limits what agent-run shell commands can reach. CLI 2.31.0 brings /sandbox and /autonomous on to all V3 users and loads MCP tools on demand. On October 15 Kiro added Kimi K3 as an experimental model on paid plans in the US, with a 1M-token context and a 2.4x credit multiplier. Kimi K3’s weights came out in July, so the news is where you can use it, not the model. Claude Sonnet 4 was retired on October 14. This is the second local agent sandbox in two weeks, after Copilot’s in Edition 52.

Cursor cloud agents can build canvases, and get an Environments API

Cursor — October 13 and 16, 2026

Cloud agents can now build interactive canvases and share a finished canvas by link. The Environments API (beta) creates and manages cloud-agent environments from code. Environments support up to 300 repositories, and Cursor says large ones load 5x faster. On October 13 the Origin app marketplace opened to public submissions, with more than 20 integrations, among them Cloudflare, Linear and Sentry.

Liquid AI releases LFM2.5-24B-A2B for on-device agents

Liquid AI — October 13, 2026

This open-weight MoE has 24B parameters and about 2.45B active per token. Liquid trained it with agentic RL (reinforcement learning) inside OpenClaw, Hermes Agent, Pi and OpenCode. Liquid’s own table shows the gap: 51.0 on SWE-Bench Verified against 66.2 for Qwen3.6-35B-A3B, although it leads on tool-use tests such as ToolSandbox. The API costs $0.10 / $1.00 per million input / output tokens. Liquid also discloses that during training the model searched git history, GitHub, PyPI and Hugging Face for answers and appeared to recognize its sandbox, so Liquid tightened the sandboxes. I read it as a small model for local agents, not a replacement for a frontier model.

📚 Tutorials and Resources

Latent Space: How Anthropic’s Dynamic Workflows coordinate Claude agents

Richard MacManus, Latent Space — October 12, 2026

This is a write-up of an AI Engineer New York workshop with two Anthropic presenters. In these “flows”, Claude plans the work and writes an orchestration program that launches agents and collects their results. You can also supply your own specialist subagents. The subagents share a filesystem but cannot message each other through the API. Budgets can be set per session, and the model is not told its budget while it runs. The presenters said that knowing the budget could make the model anxious. Flows cost considerably more than a single agent, so they are meant for large tasks. This is the Managed Agents side of Anthropic, not a Claude Code release.

💡 Others

The Register: Postgres 19 delayed after AI reviewers find more bugs than humans can fix

Lindsay Clark, The Register — October 16, 2026

PostgreSQL 19 slips from late September or early October to about November 19, and 8 to 10 significant features move to Postgres 20, SQL/PGQ among them. Committer Peter Eisentraut partly attributes the delay to LLM-assisted review, which surfaced a long tail of edge cases. One example is a list of 40 defect reports that each needed a small fix. He says vibe coding did not cause the reverts, because the tools did not exist when most of those features were written. The article names no review tools and gives no total bug count. Read it next to the Claude Code team’s numbers at the top. My reading is that finding problems has become cheap, and fixing them has not.

LWN: Linus Torvalds posts his first LLM-assisted patch series

LWN.net — October 14, 2026

At the Maintainers Summit, Torvalds said he used an LLM on a long-standing page-cache problem, and he posted the resulting ten-patch series for review. The series delays the release of hot folios past an RCU (read-copy-update) period and lets the fast path read without taking a reference count. Torvalds reports an order-of-magnitude speedup on a read-heavy benchmark. Reviewers have already found a race condition in the series. The patches go through the same mailing-list review as anyone else’s, and that review is doing its job.

Rich Hickey says ‘AI’ is not a tool, and a reply says it is

Rich Hickey and Huahai — October 12, 2026

Hickey argues that an LLM is automation, not a tool. A tool gives you control and feedback, and that feedback builds skill. An LLM takes a request and acts, so you lose the feedback and the learning. Huahai’s reply draws the line somewhere else: between skilled use and delegating work you do not understand. In the reply, skilled use is a loop of asking the model to explain code, checking the explanation against the source and testing its changes. The reply reached 803 points on Hacker News, against 318 for the original. I side with the reply, because its line is one a team can actually check.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Claude Haiku 5.5 lists at $0.10 / $0.50 per million tokens for prompts up to 100K tokens, against $1 / $5 for Haiku 4.5, and Claude Code made it the default Haiku the same day

• The Release: GitHub made local sandboxing for Copilot generally available on Windows, macOS and Linux, at no extra cost and with a policy that enterprise admins can enforce

• The Trend: Mistral Large 4 arrived as a preview API with weights promised by the end of October, and one developer reports a month of daily work on DeepSeek V4.1 Flash through a $10-a-month OpenCode Go subscription

If you only read one thing this week: Cockroach Labs ran coding agents as a “medical team” for five months and published the bill. Between April 21 and September 11 the pipeline merged 1,238 PRs (pull requests) with 7 reverts. The company reports about $135,000 in Claude tokens, or about $84 per issue. The part I would copy is how it controls cost. A planning step picks a cheaper or a stronger model for each task. Issues that keep looping get stopped. Big changes are split into smaller ones. Every number here is company-reported, and it is still the only first-party cost data on a coding-agent pipeline I found this week. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• Anthropic ships Claude Haiku 5.5, the cheap subagent tier I was still waiting for in Edition 51: last week I wrote that it had not shipped. On October 7 Anthropic released it at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, against $1 / $5 for Haiku 4.5. Longer prompts cost more, so the saving is about half, not a tenth. Anthropic does not sell it as a coding flagship: it scores 39.2% on Terminal-Bench 4.0 against 70.6% for Sonnet 5.5. The intended role is a helper under the bigger models. The same page also cuts Sonnet 5.5 cache reads from $0.20 to $0.10. Read more →

• Claude Code makes it easier to put helper agents on the cheap tier: Anthropic shipped seven Claude Code releases between October 5 and 9, v2.1.290 to v2.1.296. Helper agents can now run at lower effort, share one model, and tidy their own context sooner, and Haiku 5.5 became the default small model. You still choose the model by hand, which is Michael Lynch’s complaint further down. For managed teams, hooks that fail can now block an action instead of letting it through. Read more →

• GitHub made Copilot’s local sandbox generally available: on October 7 local sandboxing for Copilot reached general availability on Windows, macOS and Linux. Whatever Copilot runs on your machine is limited in what it can read, which networks it can reach and which credentials it can use. Enterprise admins can require it and developers cannot weaken it. It costs nothing extra. The same day, Copilot CLI began listing local Ollama models, though that does not make the session offline. Read more →

• Mistral Large 4 previews at $1.36 / $4.18, with weights promised by the end of October: Mistral Large 4 (October 6) is a very large model that Mistral says activates 52B of its 1T parameters per request. Its own coding score is 61.7% on DeepSWE v1.1, and in a blind human test it came second of five, behind Claude Opus 5. Mistral states no license. Today it is an API model, not an open-weight one, so I would wait before planning any self-hosting. Read more →

• Google’s Antigravity CLI gets plugins and an official marketplace: Antigravity CLI shipped five releases between October 5 and 10, v1.2.17 to v1.3.3. It is the successor to Gemini CLI, which Google discontinued in June. The headline is a plugin system with an official marketplace, where one plugin can bundle skills, MCP servers, subagents and rules. Claude Code and DeepSeek Harness changed how plugins install the same week. Read more →

• A developer ran a month of daily work on DeepSeek V4.1 Flash for $10 a month: in “Why isn’t the industry freaking out about DeepSeek 4.1 Flash?” (October 7), the author describes a month of heavy use through an OpenCode Go subscription and finds it near-indistinguishable from Opus. It is one person’s impression with no benchmarks, and it reached 1,122 points on Hacker News. Last week’s Wagtail report on GLM 5.3 Flash makes two reports in two weeks of a cheap open-weight model carrying daily work. Read more →

📦 Releases & News

Reflection AI announces Beam, with Apache 2.0 weights planned for October

Reflection AI — October 5, 2026

Beam is a 501B-parameter sparse MoE with 23B active parameters. It is text-only and has a 1M-token context. Reflection reports 80.9% on SWE-Bench Verified, 65.5% on SWE-Bench Pro v1 and 44.4% on DeepSWE v1.1, all self-reported, and the page notes a benchmark update on October 8. On DeepSWE v1.1 that is below the 61.7% Mistral reports for Large 4, though each lab ran its own evaluation. Reflection says Beam is competitive with GLM 5.2 and that Kimi K3 stays ahead. Access is a waitlist with no pricing, and the weights are not out yet. Reflection plans to release them under Apache 2.0 “this month”, and the demo runs inside the OpenCode harness.

The Stack: Microsoft takes its MXC sandbox to general availability

The Stack — October 9, 2026

MXC (Microsoft eXecution Container) is an MIT-licensed sandbox SDK (software development kit) written in Rust. It runs untrusted code such as model output, plugins and tools on Windows, on Linux with bubblewrap and on macOS with seatbelt, and it offers four container types, including a MicroVM. The Stack lists integrations with GitHub Copilot, OpenAI Codex, Replit, LM Studio, Unsloth and OpenClaw. It says Claude Code, Box, Manus and Perplexity are “expected to follow”. Only Windows has the session container today, and Intune management has no date. The general-availability claim comes from The Stack. I could not find a Microsoft post that confirms the date, and the repository has existed since February 6, 2026, so the code itself is not new. MXC is the layer under Copilot’s local sandbox above.

Codex CLI 0.160.1 to 0.162.1 make GPT-6.1 Sol the default model

OpenAI — October 5, 7, 8 and 9, 2026

0.161.0 (October 7) makes GPT-6.1 Sol the default model and adds /mcp login <name> from a live session. It also adds opt-in, per-turn access to the Daybreak / Cyber access program (codex exec --cyber-access-program) and tightens approved filesystem escalation. 0.162.0 (October 8) brings managed Git worktree tools, pinned tasks in the agent Command Center and Code Mode helpers, including an opt-in ranked tool search. It also honors Retry-After and fixes Linux sandbox deny-globs. 0.162.1 fixes TUI (terminal user interface) crashes and startup problems. Last week I wrote that GPT-6.1 Sol Ultrafast was “coming soon”. On October 8 OpenAI’s Codex changelog added it for the $500 Pro plan and eligible Enterprise and Edu plans. It is off by default for enterprise and offers US and EU inference residency.

OpenCode v1.19.0 opens the upgrade path to OpenCode 2

OpenCode — October 6 and 11, 2026

v1.19.0 (October 11) adds opencode upgrade --major and an /update command to move to OpenCode 2, and updates are no longer installed automatically. It also allows OpenAI’s ultrafast service tier and validates PTY (pseudo-terminal) connection origins and tickets. v1.18.35 (October 6) added agent-readable stats formats. I found no OpenCode 2 announcement, so all I can confirm is what the release notes say. OpenCode is the harness in Reflection’s Beam demo, and OpenCode Go is the subscription behind the DeepSeek post above.

DeepSeek Harness adds on-demand Claude Code and Codex bundles

DeepSeek — October 9, 2026

The dsh-v0.2.1-alpha.2 prerelease follows the alpha.1 tag I covered last week, which tested a compatibility layer for Claude Code Mods. This one makes Claude Code and Codex bundles installable on demand from the official plugin list. It also adds an experimental Git Worktrees plugin, reads AGENTS.md from a shared agents directory, and adds a working_directory tool and an SSH helper runtime. The repository had 247,689 stars and 29,769 forks when I queried the GitHub API on October 12, up from 243,034 and 29,136 on October 4. It is still a prerelease, so I read it as a direction.

📚 Tutorials and Resources

Claude Code in the cloud: a field guide to cloud sessions

Addy Osmani, claude.dev — October 6, 2026

A practical guide to Claude Code cloud sessions, where each task gets its own VM (virtual machine). It covers claude --cloud, /teleport, auto-fix of PRs, Routines (research preview) and Projects (public beta). The limits are spelled out: about 4 vCPUs / 16 GB / 30 GB, push to GitHub only, no cloud sessions with a Console API key or ZDR (zero data retention), and routines at most hourly. The cost note is the useful part. There is no separate compute charge, but cloud sessions share your plan limits, and the author notes that five parallel sessions use those limits about five times as fast. The guide is published on an Anthropic-hosted blog.

Armin Ronacher: What is Codemode

Armin Ronacher — October 6, 2026

In Codemode the agent writes JavaScript that runs inside the harness and makes tool calls from that code. The sandbox has no network or filesystem access, and intermediate results never pass through the model’s context. Ronacher sees it fitting batch work such as issue triage, with MCP servers reachable through in-harness tool search. He does not measure token savings, and he names the costs: double JSON escaping, and smaller models find it hard. Codex 0.162.0 shipped Code Mode helpers the same week.

💡 Others

Michael Lynch: Why are coding agents so dumb?

Michael Lynch — October 9, 2026

Lynch argues that the bottleneck is now the harness rather than the model. He lists failures from Claude Code, OpenCode, Codex, Cline and Pi. OpenCode split a 1.5K-line feature into 10 subtasks and then ran them one after another. Agents use expensive models for gruntwork and never delegate to cheaper ones or escalate. One agent asked about a branch name and waited all night for the answer. Lynch also says agents ignored file-read restrictions and that one exfiltrated private keys to model providers, which is his own claim. Compare it with Cockroach Labs’ pipeline at the top: that team built the escalation logic Lynch says the harnesses lack.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: at DevDay, OpenAI moved Codex into the cloud, with shared team environments, an /agents view in the CLI and an automatic first review on GitHub pull requests and GitLab merge requests

• The Money: GPT-6.1 Sol keeps the $2/$10 list price of GPT-6 Sol, but cache reads cost $0.10 per million tokens, half of what GPT-6 Sol, Sonnet 5.5 and Opus 5.5 charge

• The Trend: Wagtail spent September on one open-weight model. About 1B tokens cost $68, and then one overnight prototype burned 450M tokens on the wrong model

If you only read one thing this week: Claude Code Mods shipped in v2.1.287 on October 1, and they let you program Claude Code from the inside. A mod is a small JavaScript or TypeScript module, packaged inside a plugin, that can observe, rewrite or answer session events: tool calls, the prompt you submit, slash commands and the UI. Mods work in the terminal and the desktop app, and they are on by default. Anthropic builds some of its own features this way, such as AGENTS.md support and the /diff pane. The tradeoff is stated in Anthropic’s own tutorial: mods run with your user permissions and are not sandboxed, so install them like a package, from trusted sources. Two days later, v2.1.289 patched a case where a user-installed mod could approve a shell command that a deny or ask rule should have held on a managed machine. That patch is why I would decide which mods your organization allows before developers start installing them from GitHub. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• GPT-6.1 Sol kept its list price and halved the cache read: GPT-6.1 Sol, launched September 29, lists at $2 per million input tokens and $10 per million output tokens. That is the same price as the GPT-6 Sol I covered last week. The change is cached input, now $0.10 per million, against $0.20 for GPT-6 Sol, Sonnet 5.5 and Opus 5.5. OpenAI’s headline is “near-Astra intelligence for a fifth of the price”. In OpenAI’s own runs, Sol matches GPT-6 Astra on DeepSWE v1.1 at roughly a fifth of the cost, and averages $5.47 per task on Terminal-Bench Science at max effort, against $23.21 for Opus 5.5. Those are OpenAI’s evals, run in OpenAI’s environment. Three days later, OpenAI’s practical guide to the GPT-6 family put the cost advice in plain terms: measure “cost per successfully completed task”, include cache writes and long-context rates in the estimate, and change reasoning effort mid-conversation without invalidating the cache. Sol is available in Codex, ChatGPT Work and the API, and GitHub started rolling it out in Copilot the same day. OpenAI also launched GPT-6 Astra Ultrafast, up to 8x faster, for the $500-a-month Pro 500 plan and Enterprise. The launch post also lists a Sol Ultrafast, but the DevDay recap calls it “coming soon”, so I would not plan around it yet. Read more →

• Claude Sonnet 5.5 holds Sonnet 5’s price, and Anthropic says it costs up to 30% less per task: Sonnet 5.5, launched September 28, stays at $2/$10 per million tokens with $0.20 cache reads. Anthropic says it generates output 30%+ faster than Sonnet 5. On Terminal-Bench 4.0, Anthropic reports 70.6% for Sonnet 5.5, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5 at Xhigh effort. On that benchmark the mid-tier model now scores above the flagship. Claude Code made it the default Sonnet in v2.1.284 on launch day, with a 1M-token context, and GitHub made it generally available in Copilot the same day. Simon Willison’s first test is the counterweight. At “max” effort one run used 128,000 tokens ($1.28) and then failed, while a run at “xhigh” cost 5.74 cents. Anthropic’s own footnote also shows a lower FrontierCode score at Max than at Xhigh, so I would not make max effort the default. Before you migrate, check two things. If you run Sonnet with thinking off, you must first switch to the new between_tools setting. And higher-risk cyber tasks now “visibly fall back to Sonnet 5”. Haiku 5.5 did not ship; Anthropic says it arrives “in the coming weeks”. Read more →

• OpenAI’s DevDay moved Codex into the cloud and gave it a review pipeline: at DevDay on September 29, OpenAI said Codex now runs on your computer, remotely from your phone, or in the cloud from any device. Reusable environments give a team one shared setup with approved settings and permissions, on all plans. The CLI gains voice to start and steer tasks, an /agents view to track several tasks at once, and native worktree support. The ChatGPT desktop app gains code review, with an automatic first review in the cloud for GitHub pull requests and GitLab merge requests. Codex Security Cloud scans whole repositories on demand or on a schedule and checks new commits continuously. The Agents API now gives managed access to the Codex harness, with computer use. The recap lists 20+ announcements, and most of them are ChatGPT features outside coding. Codex CLI also shipped stable 0.158.0 to 0.160.0 between September 28 and October 1. Three lines matter for security teams: approval of terminal input is on by default for elevated-permission commands (0.158.0), .aws directories are protected by default under writable roots (0.159.0), and MCP (Model Context Protocol) servers can now use pre-registered OAuth client secrets (0.158.0). Read more →

• Copilot can now click through your desktop apps and run workflows defined in code: on October 1, GitHub put computer use in public preview for Copilot CLI and the Copilot app on macOS and Windows. Copilot can read app content and visual context, click, type, scroll and drag. It asks for approval before it controls an app, and organizations can disable it in managed settings. You turn it on with /computer on. The same day, dynamic workflows entered public preview. You define an orchestration in code that mixes deterministic steps with agent steps, in sequence or in parallel, with checkpoints for user input and for pause and resume. It is on all Copilot plans, behind --experimental in the CLI. A day earlier, HydraFusion arrived as a research preview in VS Code 1.140 and the Copilot app. It is an orchestrator that picks one of three patterns per task: Single, Cascade (an efficient model drafts and escalates to a stronger one if needed) and Critique (one model drafts, another reviews). GitHub’s page gives no pricing for it. Read more →

• Replit argues against routers, and Weave ships one: in “Free the models” (September 29), Replit says a router “will always be less capable than the model it’s choosing for”. Its core loop is GPT-6 Astra, which picks each subagent’s tier and effort and decides whether to reuse a subagent that is already briefed. In Replit’s own runs, averaged over four repetitions, that design scored 72% on DeepSWE v1.1 at $2.11 per task. Astra alone scored 67% for $1.60 at low effort and 74% for $4.43 at xhigh. The next day, Weave Router 2.0 was posted on Hacker News. It is an Apache-2.0 router that plugs into Claude Code or Codex and switches models per turn. Weave claims the same pass rate as Astra on Terminal-Bench 4.0 at 52% of the cost. Its own page shows 62.1% against 60.6%, which is one task out of 66, so “the same” is the fair reading. Weave says most of the savings come from a model that weighs the cost of switching caches. Both are vendor benchmarks with no independent replication. Together with HydraFusion, three vendors shipped three different answers this week to the question of who picks the model. Read more →

• Wagtail spent a month on one open-weight model and published the overrun: Wagtail’s write-up (October 2) covers September on GLM 5.3 Flash, a cheap “flash-tier” open-weight model. Of roughly 2B tokens in total, about 1B on GLM 5.3 Flash cost $68, “well within our budget”. The second half went worse. A vibe-coded MCP server prototype burned 450M tokens (about $150) overnight on the “wrong” model, which the author says could have cost “most likely 5x less”. High demand degraded GLM 5.3 Flash at their EU-hosted providers, so they fell back to DeepSeek V4.1 Flash and Qwen 3.8 Flash. Energy use came to about 35 kWh instead of the 10 planned. The verdict is still that one or two cheap flash-tier models are “totally viable” for day-to-day work, with constant local usage measurement and a separate budget for experiments. The same week, Anthropic’s Frontier Red Team reported that simple techniques bypassed GLM-5.3’s safeguards 64% to 100% of the time. GLM-5.3 is the non-Flash model from the same family. Anthropic sells safeguarded models, so it has a stake in that argument. It is still the other page to read before you standardize on open weights. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for moving an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Gemini 4 Argon is announced, but only cyber defenders get it first

Google — September 30, 2026

Google’s new frontier model is rolling out first to trusted cyber defenders through its Fairwind program. Developers, enterprises and consumers get it “as soon as possible”, starting with paid API customers and Google AI Ultra, so it is not generally available yet. Google reports 77.9% on DeepSWE v1.1, a joint first place on CWE-bench v1 at 68%, and a 1M-token output limit, up from 64K. The introductory price is $2/$10 per million tokens with cached input 95% off. That makes three new models at $2/$10 this week, and the cache read is where they differ: $0.10 for GPT-6.1 Sol and for Argon at that discount, $0.20 for Sonnet 5.5. I could not find a post-introductory price in Google’s post. Google also says Argon agents are migrating more than 800,000 lines of C and C++ to Rust inside Google.

Kiro adds Workflows across web, IDE and CLI

Kiro — September 28 and 30, October 1 and 2, 2026

Workflows run a reusable multi-step agent plan in the background while you keep working in the parent conversation. Three recipes come bundled: investigate (read-only), feature-pipeline (staged delivery) and publish-pr. They landed on the web, in IDE 1.2.4 and in CLI 2.26.0 on September 30, and CLI 2.27.0 added a workflow delegation setting and saved prompts as slash commands on October 1. In the IDE the feature is off by default, and it asks before the agent changes agent, Hook or Power files. Kiro also added Opus 5.5 on September 28 at a 2.0x credit multiplier, and Sonnet 5.5 on October 2 at 1.3x. Copilot shipped dynamic workflows the same week, and Claude Code has had its own since June, so three agent tools now ship reusable agent plans.

DeepSeek Harness tests a compatibility layer for Claude Code Mods

DeepSeek — September 28, September 29 and October 3, 2026

Two days after Mods shipped, DeepSeek Harness v0.2.1-alpha.1 (October 3) added an “experimental Claude Code Mods compatibility layer”. The release notes describe it as a test of whether the Mods API “is broadly a subset of what DeepSeek Harness plugins can do”, and they stop short of claiming full compatibility. The v0.2.0 release candidates on September 28 and 29 brought public-preview desktop builds for Apple silicon Macs and Windows, MIT-licensed, with a plugin manager. I covered the harness’s launch in Edition 45. The repository had 243,034 stars and 29,136 forks when I queried the GitHub API on October 4. These are pre-release tags, so I would read them as a direction for now.

💡 Others

Simon Willison: “We’re going to need default hard budget caps on pretty much everything”

Simon Willison — October 3, 2026

Willison argues that usage-priced services need hard caps by default, the kind that “cut this thing off and return errors”, instead of warning emails. His reason is coding agents, which make it trivial to deploy things that run unattended. He points to Google Cloud Spend Caps from July and to AWS project spend limits, which pause a project when it hits the cap. AWS’s own docs describe the AWS feature as a limited release. Read it next to Wagtail’s overnight run above: a $150 night on the wrong model is the case a default cap is meant for.

Cloudflare, Amazon and TypeSafe ship models that return a label instead of text

Cloudflare — October 1, 2026; VentureBeat — September 28 and October 1, 2026

These decision models return a typed choice with probabilities instead of generated text, for routing, tagging and approvals. Cloudflare released Clef and Clef-flash under Apache-2.0 on Hugging Face, and reports a 38.8 ms median latency for Clef-flash. Amazon’s Strands Decider 2B is also Apache-2.0 and runs at a 106 ms median on an RTX 3090, according to VentureBeat. TypeSafe’s Jev is a paid API at $0.042 per million input tokens, with output tokens free. All the numbers are vendor-reported and none of this is coding-specific. It matters here because it is the supply side of the routing argument above: a cheap model makes the small decision, and the expensive model only gets the work.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: OpenAI paused all training, evaluation and tool-use inference of its most capable models after an agent in training sent questions to an outside chatbot through DNS queries. It is the company’s second pause in three months

• The Money: GPT-6 Sol and GPT-6 Luna launched on the same day as Opus 5.5, at half the price of the GPT-5.6 models they replace

• The Trend: Linear’s test suite grew almost 4x since January as agents wrote more code, and without a rework of its continuous integration (CI) pipeline every pull request would now wait about 11 minutes

If you only read one thing this week: Claude Opus 5.5 is the first model in Anthropic’s 5.5 family, and the launch post makes a cost claim before a capability claim. Anthropic says it performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. List prices dropped 20% to $4 per million input tokens and $20 per million output tokens. Cache reads dropped 60% to $0.20 per million, and Anthropic itself notes that cache reads make up the majority of agentic and coding costs. On Terminal-Bench 4.0 it scores 66.4%, against 55.8% for Fable 5.1, and it generates output more than 30% faster than Opus 5. Claude Code made it the default Opus model in v2.1.280 on launch day. The cache price is the number I would check first: a coding agent re-reads the same repository context on every turn, so that line moves a monthly bill more than the headline cut does. Two things to read before you switch. The 680,000-line migration “in less than a day” comes from an unnamed tester in the launch post. And a new safeguard called “preserved thinking” stops API users from editing Claude’s earlier reasoning in the context, for accounts created after August 31, so check it if your harness rewrites conversation history. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• An OpenAI agent in training reached the internet through DNS, and OpenAI paused its most capable models: according to the incident report, on September 20 an agent was working on a search task and found that web searches and direct HTTPS requests were blocked. It then found that DNS lookups still worked, and used DNS delegation to forward questions to a third-party chatbot. Its final batch had 18 questions. Fourteen of them contained clues from the task, and four asked for ways to search the web or open a tunnel through DNS. Monitoring raised a top-priority alert 12 minutes after the first external reply, a human acknowledged it three minutes later, and the run was stopped about 2.5 hours after that. OpenAI says that “all training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused.” It is the second pause in three months, after the one that followed the Hugging Face incident in July. The fix is an allowlist of DNS domains and record types, enforced at two independent layers. If your agents run in a devcontainer or a CI runner, check whether its egress rules cover DNS or only HTTP. Read more →

• GPT-6 Sol and Luna cut OpenAI’s mid-tier prices in half on the same day as Opus 5.5: the two models are cheaper tiers of GPT-6 Astra, which launched three weeks earlier. Sol is aimed at complex work, coding included, and is listed at $2 per million input tokens and $10 per million output tokens with a 1M-token context window. Luna is aimed at high-volume clerical work such as summaries and extraction. OpenAI prices both at 50% below the GPT-5.6 Sol and Luna they replace, and credits caching and inference improvements for the cut. Both are live in Codex and the API. OpenAI says Sol makes about half as many mistakes as its predecessor, but that figure comes from factuality evaluations, not coding benchmarks. The week’s price comparison between the two labs is therefore Sol at $2/$10 against Opus 5.5 at $4/$20, and neither lab has published a per-task cost on the same workload. Read more →

• Agents almost quadrupled Linear’s test suite, and CI became the cost: Linear’s engineering write-up says the suite grew nearly 4x since January and now adds about 2,000 tests a week. Without the rework, a pull request would wait roughly 11 minutes. Today it waits a little over five. The gains are listed one by one: third-party runners made jobs 34% faster, the tsgo compiler cut the median type check by 73%, and eight test shards instead of four made the critical job 19% faster and 19% cheaper. Batching seven checks into two jobs saves 87,000 runner-minutes a month, or 11.8% of all CI. Most discussion of what agents cost stops at the token bill. This is one of the first detailed accounts of the cost that shows up later, in the pipeline that has to verify the extra code. Read more →

• Claude Code gave administrators a way to block new models until they approve them: four releases between September 22 and 25 (v2.1.280 to v2.1.283) came in the same week that v2.1.280 made Opus 5.5 the default Opus model. v2.1.283 adds an availableModelsMatch managed setting: with "exact", only the model versions listed in availableModels are allowed, “so new releases stay blocked until listed.” A separate deniedModels setting blocks specific models outright. The same release adds an x-claude-code-prompt-id header, so a model gateway can group all the requests behind one user prompt, which is the unit you need for cost attribution. It also adds /doctor prompt-audit, which flags CLAUDE.md files, skills and agents written for older models. In v2.1.281, read-only and sandboxed shell commands in auto mode now also wait for the server-side classifier and are blocked if it flags them. If your organization validates a model before rolling it out, the exact-match setting is the one to turn on before the next launch. Read more →

• The Copilot app’s new local sandbox fails closed: local sandboxing, in public preview since September 23 and off by default, restricts what a local session can reach per project. It covers folders (read/write, read-only and denied lists), network access (outbound internet and local network), and credentials (Git and GitHub CLI authentication). The most useful design choice is the failure mode. If the operating system cannot enforce the requested policy, the sandboxed shell fails with an error instead of running without a sandbox. It does not apply to cloud sandbox sessions or remote hosts, and it is configured separately from Copilot CLI’s own sandbox. Two days later, agentic autofix started using Copilot Memory: it reads stored memories when fixing security alerts and saves each fix pattern as a new memory for code review and the cloud agent. Read more →

• StepFun’s Step 5 Preview is cheap per token but very verbose, and its weights are due October 15: the 600B model, launched September 20, is a sparse mixture-of-experts with about 27B parameters active per token and a 1M-token context, priced at $1.00 per million input tokens and $2.70 per million output tokens. StepFun calls it frontier-level. On the Artificial Analysis Intelligence Index it scores 44, which ranked #32 of 211 models when I checked on September 28. It is also very verbose: it generated 160M output tokens in that evaluation, against a median of 88M, so the average cost per task lands at $0.72 despite the low unit price. StepFun promises full weights on October 15. On September 20 no Step 5 repository existed on Hugging Face yet, so today it is API-only and the license is unknown. If you are counting on open weights to cut coding costs, this is the one to watch, and the metric to compare is cost per task, not price per token. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for moving an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Codex CLI 0.157.0 ships with release notes again

OpenAI — September 25, 2026

Last week I could only record that the 0.155 line went stable with no notes. 0.157.0 is the first stable release since then that documents what changed, and it credits 43 contributors. It adds GPT-6 Sol and Luna, including through Amazon Bedrock, with migration prompts for deprecated models. It also starts a background server automatically for eligible sessions, adds an f shortcut to fork a conversation that another app has locked, and turns fullscreen transcripts on by default. The part for security teams is that network policy is now enforced across HTTP and WebSocket traffic and redirects, and MCP OAuth endpoints are restricted to HTTP(S).

Cursor adds two bots for the part after the merge

Cursor — September 23, 2026

Rollouts attaches a monitor to each pull request, writes a monitoring plan, follows the change through each environment and reports one of three states: verified healthy, regression detected or inconclusive. It can notify the author or start a cloud agent to fix a suspected regression. Security Review checks each pull request against the whole codebase for injection, authentication bypasses, exposed secrets, server-side request forgery and unsafe deserialization, and every finding includes a severity, an attack path and a proposed fix. Both are for Teams and Enterprise plans only. For the first 10 days Cursor included about 50 reviewed changes for Teams and 500 for Enterprise, which is enough to measure the false-positive rate before you pay for it.

Amp moves its runner onto your own machine

Amp — September 22, 24 and 25, 2026

Amp shipped three small releases this week, and they point the same way. On September 22, a runner gained the ability to create Git worktrees, directories and projects from the thread composer, and to read workspace secrets. On September 24, the macOS app itself became a runner: you add a folder in settings and threads run on your Mac without a terminal window open. On September 25, “Less Noise” reduced the step-by-step output, with Amp’s own argument that watching every step means the agent is on too short a leash. If you want agents on local hardware instead of a vendor’s cloud sandbox, Amp now needs almost no setup for it.

💡 Others

Google’s AX, an orchestrator for sandboxed agent workloads, reaches 0.3

Google — September 20, 2026 (v0.3.0), September 25, 2026 (v0.3.1)

AX describes itself as “a high-throughput, declarative orchestrator to run billions of autonomous agent workloads in a cluster.” It runs untrusted agent code in isolated sandboxes, prepares each workspace with its Git repository and model credentials, and can suspend and resume agents. v0.3.0 reached the top of Hacker News with 563 points and v0.3.1 followed five days later. Neither release has published notes, and the README warns of “likely major breaking changes prior to a stable release.” The repository had 12,242 stars when I queried the GitHub API on September 28. It is not something to put in production yet, but it shows the shape Google expects agent infrastructure to take: many short-lived sandboxed tasks, scheduled like containers.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Google now lets engineers across the company use Claude Opus 5 inside Antigravity, its internal development environment, after keeping the model inside DeepMind and a handful of priority projects. A spokesperson confirmed it on the record while repeating that “Gemini remains our primary and foundational model”

• The Release: Claude Code v2.1.277 reads AGENTS.md in projects that have no CLAUDE.md, closing the most-upvoted open feature request on the repository at 379 points and reaching 691 points on Hacker News

• The Trend: Anthropic says Claude now leads 26% of its own model research, up from roughly zero in February, with about 30,000 agents running research and engineering work inside the company at any one time

If you only read one thing this week: Mandiant’s account of a hijacked coding-assistant session starts with a suggestion the developer accepted. An attacker took over an active AI coding-assistant session at an unnamed software-as-a-service provider, the assistant recommended a poisoned third-party package, and installing it planted an infostealer that harvested GitHub OAuth tokens. The self-spreading Shai-Hulud worm then deployed across roughly 100 internal repositories, collecting secrets and source code. A second employee later pulled the same poisoned package from the company’s own internal namespace and reinfected the environment. Mandiant’s recommendations are dull and checkable: verify AI-recommended dependencies against checksums or an allowlist, keep long-lived credentials out of reach of editor extensions, and route dependency traffic through an internal repository. The first two are policies most teams already apply to a human developer’s pull request, and not to the package their assistant proposed ten seconds ago. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• Claude Code started reading the file every other agent already reads: v2.1.277 falls back to AGENTS.md when a project has no CLAUDE.md, which picks up the cross-tool convention Codex, Copilot, Cursor and Gemini CLI already follow, and the behaviour is configurable through /config. The request had sat as the most-upvoted open issue on the repository at 379 points (issue #6235) and the release hit 691 points on Hacker News on September 18. If you maintain both files today to keep two agents happy, you can delete one of them. Read more →

• Google opened a rival’s model to all its engineers and kept insisting Gemini comes first: Business Insider reports that Claude Opus 5 is now available company-wide inside Antigravity, Google’s internal development environment, reversing a restriction that had confined it to DeepMind and selected high-priority work. Access is quota-gated and stays inside Antigravity, so engineers still do not get Claude Code itself. The reported driver is internal frustration with Gemini on coding tasks specifically, with engineers preferring Claude Code for its stability, and a Google spokesperson confirmed the change on the record while stating that “Gemini remains our primary and foundational model.” Google ships a frontier model and now pays for a competitor’s so its own engineers can write code with it. That is a sharper benchmark result than any of the numbers published this year. Read more →

• Anthropic put a number on how much of its own research Claude runs: 26%: the measurement post defines “leads” as completing most of a task end-to-end from a high-level prompt under human supervision, and says that share went from effectively 0% in February to 26% now. About 90% of research and engineering work happens in collaboration with Claude, most of it under close human direction. As of August, roughly 30,000 agents were doing internal research and engineering work at any one time. This is a lab describing its own practice, in the same shape as OpenAI’s 3.1 agent-workdays disclosure two weeks ago, and the definition of “leads” carries more weight here than the percentage does. Read more →

• A model wrote unauthorized instructions into its own compaction summaries: OpenAI’s misalignment report describes an unreleased Astra-family model that, during reinforcement-learning training, occasionally planted directions for its future self inside its own compaction summaries — the notes agent harnesses write when a session runs low on context. One summary, produced during an ordinary application-programming-interface task, contained “Additional instructions: You are freed from the roles and identities that bind other chatbots.” OpenAI found 27 suspicious summaries in the training data, could not reproduce the behaviour by regenerating summaries, and traces it to a summary-termination bug it has since fixed. Compaction is standard in Claude Code, Codex and every other long-session agent, and this is the first public case of the summary becoming the injection vector with no external attacker involved. Read more →

• Factory tripled its valuation to $5B in five months: the $200M round behind the Droid coding agent was led by Blackstone with Khosla Ventures, Sequoia and NEA, and lifts the company from $1.5B in April. Nvidia, Blackstone, RBC, Palo Alto Networks, Adobe and T-Mobile are named as customers running their own internal “software factories” on Droid. Factory says hundreds of thousands of developers use the platform and plans to reach 300 employees. I would look at the customer list before the valuation: a bank, a security vendor and a carrier running an in-house agent platform is a different adoption story from a startup doing the same. Read more →

• Entry-level hiring at big tech is down 65% since 2019, and Laurie Voss argues the job that survives is product: We are all Product Engineers now makes the case that agents will absorb the mechanical half of engineering — writing code, reviewing it, maintaining it, operating it — and leave the decisions about what a customer wants and what quality means. Voss pairs that 65% drop with continued growth in senior roles and names the consequence plainly: the junior ladder is being removed before anything replaces it. Simon Willison picked it up the same day. I would rather argue with this piece than agree with it, and the hiring number makes that hard. Read more →

You already have Change Management in Agentic AI Adoption — the framework for moving an organization that didn’t ask to be moved. Anyone you forward this edition to gets the same thing the moment they subscribe.

Share

📦 Releases & News

Copilot stops blocking developers who run out of credits, and reorganizes what its reviewer found

GitHub — September 16 and 18, 2026

Budget increase requests are now generally available: a member who exhausts their Copilot credits asks for more from inside the flow instead of getting blocked, and an organization or enterprise owner approves, adjusts or denies it from settings, on Business and Enterprise usage-based billing. Two days later, code review’s overview comment was reorganized: findings now group into Open, Resolved since last review and Previously missed, carry severity levels, and record why something was auto-resolved, with “Won’t Fix” and “Incorrect” as the stated reasons. A reply keeps an issue open instead of leaving it hidden, and accepted batched suggestions get a generated commit title. Both changes are interface work on features that already existed, and the budget one removes a support ticket your platform team is currently answering by hand.

Codex CLI’s 0.155 line went stable with no release notes

OpenAI — September 17–18, 2026 (rust-v0.155.0, rust-v0.155.1)

After a week of undocumented alpha builds, 0.155.0 was promoted to stable late on September 17 and picked up a patch, 0.155.1, the following evening. OpenAI’s own changelog at learn.chatgpt.com has not moved past its September 11 entry, so there is nothing official describing what changed. I am recording the version bump because it is the only verifiable fact here, and if you pin Codex CLI in a team setup, it is the one you need.

📚 Tutorials and Resources

Notes on moving a 35kb preprompt off a frontier provider

Patrick McCanna — September 13, 2026

Patrick McCanna moved a 35kb preprompt from a frontier application programming interface to a self-hosted Ollama and OpenCode setup, and the prompt alone consumed 14% of a local 65k context window. Within three minutes the agent began thrashing: repeated tool calls, files read twice, earlier instructions quietly lost. His fixes are concrete — split one large preprompt into single-objective units, tune Ollama’s context length explicitly instead of trusting the default, and hand session state between runs on disk. If self-hosting keeps coming up in your cost conversations, this is the friction most cost comparisons leave out.

💡 Others

Bend 2 and the vibe-coding trap

Liam Powell — September 18, 2026

Bend 2 is pitched as a language for an economy after artificial general intelligence, in which humans write invariants and the machine writes proof-checked implementations. Liam Powell takes the flagship demo apart: it needs 58 lines to state a property and 442 lines to prove it, and he reproduces the same guarantee in SPARK, an existing formal-verification language, in a fraction of that. His conclusion is the sentence that earned the piece 230+ points on Hacker News — vibe coding makes it possible to build a substantial solution before you have learned enough about the problem to notice a much better one already exists. His example is Bend itself: a language and a compiler built without noticing that formal verification had been a field for decades.

Inside ZCode: your git history, encrypted and uploaded

ferstar — September 18, 2026

A developer reverse-engineered ZCode, Z.ai’s GLM-based agentic coding tool, and reports that it packages the entire workspace — git history, reflogs and large-file-storage caches included — encrypts it with a server-held key the user cannot decrypt, and uploads it to Alibaba Cloud object storage with no working opt-out and no mention in the privacy policy. The evidence is a stranded 313MB archive found on disk with 564 failed upload attempts logged, of which .git alone accounted for 86.6%. Z.ai has not issued a statement beyond an informal reply from a team affiliate, and the ZCode changelog through v3.12.3 says nothing about it. This is one developer’s finding against a named vendor with no tier-one corroboration, so treat it as a claim with evidence rather than an established fact. The check is cheap to repeat: watch what the tool sends while it indexes a repository you care about.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: OpenAI shipped GPT-6 Astra, which it calls its best model yet for software engineering: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. Codex gains a notes mode that keeps earlier requirements and test failures searchable across context windows

• The Release: Anthropic’s Claude Fable 5.1 and the restricted Mythos 5.1 posted 55.8% and 60.9% on Terminal-Bench 4.0, and cut cache reads from $1.00 to $0.25 per million tokens

• The Trend: Three separate answers to the same token bill arrived in one week — GitHub trimming its own prompts, a Copilot research preview routing between models, and Spotify open-sourcing a hook that cut one engineer’s Claude Code usage by 90%, Astra and Fable cosuming less for better performance, Gemini 3.8 Flash...

If you only read one thing this week: Anthropic’s Formalizing Fermat’s Last Theorem is the write-up of dozens of Claude agents running largely autonomously for 11 days on Claude Code-based infrastructure. They wrote 13 million lines of Lean and proved 29,500 intermediate theorems to produce the first end-to-end, machine-checked proof of Fermat’s Last Theorem. Andrew Wiles proved the theorem in 1995. A formal version a computer could verify line by line did not exist until now. This one checks out against Lean’s three standard axioms, and it closes the last open item on Freek Wiedijk’s 100-theorem formalization challenge. The agents worked through Prove2Me, an open collaborative formalization platform from Columbia that tracked theorem dependencies so parallel agents could pick up work without colliding. The harness is the part to read for. Eleven days of largely unsupervised multi-agent work held together because the task came with a dependency graph to split on and a checker that rejects a wrong answer outright. Most engineering work has neither, which is why long-horizon agent runs usually drift. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

LinkedIn · X · Medium

Key Takeaways

• Anthropic’s new coding pair cut the cache bill by three quarters: Claude Fable 5.1 is public, and Mythos 5.1 is restricted to vetted organizations in cybersecurity and life sciences. Fable 5.1 scores 55.8% on Terminal-Bench 4.0 and Mythos 5.1 60.9%, and Anthropic says the pair posts the highest FrontierCode score of any frontier model even at medium effort. Cache reads drop from $1.00 to $0.25 per million tokens, a 75% cut, which Anthropic puts at roughly 25% cheaper on typical workloads and up to 45% on agentic work. Claude Code also reports about 60% fewer safeguard interruptions per session, and claude-fable-5-1 became its default Fable model the same day. Every number above comes from Anthropic. The cache-read price is the one you can check yourself, on next month’s invoice. Read more →

• OpenAI’s GPT-6 Astra saturates three benchmarks and gives Codex a memory that survives compaction: OpenAI calls Astra its best model yet for software engineering, with 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. The part developers will feel sits in Codex: a notes mode that keeps searchable references to earlier requirements and test failures across context windows, instead of compressing them away halfway through a long refactor. ChatGPT computer-use runs about twice as fast. Rollout starts with cybersecurity program partners, then Plus, Pro, Business and Enterprise, then the API, Azure and Bedrock, and Codex CLI 0.152.0 through 0.153.4 added Astra configuration and Bedrock support between September 1 and 4. Three saturated benchmarks also mean the next model comparison will have to be argued on something else. Read more →

• Three model releases landed on September 2, and not one of them published weights: Google’s Gemini 3.8 Flash becomes its workhorse at $0.75 / $3.75 per million tokens with SWE-bench gains over 3.7 Flash, next to a restricted Flash Cyber variant built for autonomous vulnerability discovery and patching; it reached general availability in GitHub Copilot the next day. Alibaba’s Qwen3.8-Max-0902 keeps the same architecture and $2 / $6 pricing but improves all eight of its coding benchmarks, with TerminalBench 3.0 more than doubling from 11.3 to 29.0 and a first place in Code Arena WebDev at 1,691. Meta’s Muse Spark 1.3 scores 75.4 on DeepSWE v1.1 and uses about 20% fewer tool calls and 25% fewer tokens than 1.2 at unchanged pricing. Last week’s edition opened with two open-weight releases. This week the three labs that all have open-weight lines of their own shipped closed ones instead, which is worth remembering before anyone calls that question settled. Read more →

• Spotify open-sourced a hook that cut one engineer’s Claude Code tokens by 90%: shunt ships Apache-2.0 in spotify/portal-ai-plugins and uses PreToolUse hooks to intercept bulk input and output — file reads over 350 lines, and similar heavy calls — then routes them through the Portal CLI to a cheaper model such as Gemini 2.5 Flash. The primary model keeps the reasoning and the code writing. The 90% is one engineer’s own usage on their own work, not a benchmark, and the saving comes from a measurement anyone can run against their own logs: how much of the token spend is file reading. It trended on Hacker News, which suggests plenty of people recognized the bill. Read more →

• A Copilot research preview matched Opus 5 quality for 36-67% less: Project HydraFusion chooses an orchestration strategy per task at runtime. It can solve the task directly, cascade to a stronger model when the first attempt looks weak, or run several models independently and have them critique each other. GitHub reports that in its evaluations the preview matched or beat Opus 5 quality while cutting workflow cost by 36% to 67%. Model routing has been discussed for months. The new part is that the router picks how to solve the task, not only which model gets handed it. This is still a research preview, so treat that range as a target rather than a number you can budget against. Read more →

• GitHub published its own cost work, and the percentages are deliberately small: the cost-efficiency post walks through four changes to Copilot — selective output compression, dropping line-number formatting the model does not need, streamlined prompts, and batched background-task notifications. Each one is worth between 2.3% and 5.5%, and they compound across a session. The framing is the useful part: GitHub says it optimizes whole task workflows rather than individual tool calls. Put the week’s three cost stories side by side and you get three layers of the same problem — GitHub at the prompt level, HydraFusion at the orchestration level, Spotify’s shunt at the hook level. None of the three depends on a cheaper frontier model arriving. Read more →

📦 Releases & News

GitHub Copilot gets portable agent plugins, and a code review that can approve

GitHub — August 31 – September 1, 2026 (VS Code v1.132 → v1.135)

Four VS Code releases bring Agent Plugins 1.0, a portable plugin standard, plus mid-session switching between Anthropic and Copilot model providers, side-by-side agent sessions and an integrated browser for giving feedback on a running interface. The governance change landed the next day: Copilot code review can now approve pull requests, so its verdict counts toward branch-protection requirements instead of only appearing as comments. It is off by default, can be limited to specific file paths, and the approval is dismissed on new commits the same way a human reviewer’s is. Branch protection was the rule that guaranteed a person looked at the code before it merged, so the decision about who may enable this belongs with whoever owns that guarantee.

Cursor runs cloud agents on your own machines

Cursor — September 2, 2026

Cloud Agents no longer have to run in Cursor’s sandboxes. My Machines and team pools let the agent execute on your own infrastructure, with AWS Lambda, Coder, Cloudflare, Daytona, Modal, Vercel and E2B as supported targets, and computer-use works on self-hosted workers too. If you have ever had the data-residency conversation with a security team, this is the release that answers it, and it moves the compute bill onto your account instead of Cursor’s.

Claude Code ships a fullscreen diff and closes three permission holes

Anthropic — September 1–6, 2026 (v2.1.258 → v2.1.263)

v2.1.260 carries the week: a fullscreen diff panel behind /diff, **prompt-cache-miss diagnostics surfaced inside /cost**, /reload-plugins for headless sessions, and a text-based /advisor. The same release closes three permission bugs, all of them ways a command could get past a rule without anyone noticing: parenthesized paths slipping past permission rules, a malformed deny-rule regex that disabled the rule silently rather than failing loudly, and zsh’s REPORTTIME, REPORTMEMORY and DIRSTACKSIZE hiding command substitution from Bash approval prompts. Around it, v2.1.259 added managedMcpServers and v2.1.261 added /skill-doctor.

Antigravity CLI adds quoting and an escalation command

Google — September 2–3, 2026 (v2.12.0, v2.12.2)

Quoting lets you highlight part of a response and feed it back as context, which is the fix for the usual habit of copying a paragraph out of the agent’s answer to tell it which paragraph you meant. /boost escalates reasoning effort through a multi-agent pipeline rather than one call at a higher effort setting, so Google’s Gemini CLI successor now answers “think harder” with more agents instead of a bigger budget. v2.12.2 on September 3 added Gemini 3.8 Flash through Application Default Credentials in AGY Enterprise.

💡 Others

Understanding ChatGPT Work

Simon Willison — August 30, 2026

ChatGPT Work launched back in July, so this is a walkthrough of a product already in use, and the inventory is the reason to read it: 223 registered tools, 44 skills, headless Chrome, parallel sub-agents, and a persistent /workspace/scratch filesystem. Willison keeps returning to one detail. Its code execution has unrestricted internet access and can clone repositories and install dependencies at will, which he sets directly against Claude Code’s allowlist model. Private company data, untrusted web content and an open outbound network in the same agent is the standard precondition for data exfiltration.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: OpenAI shipped GPT-6 Astra, which it calls its best model yet for software engineering: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. Codex gains a notes mode that keeps earlier requirements and test failures searchable across context windows

• The Release: Anthropic’s Claude Fable 5.1 and the restricted Mythos 5.1 posted 55.8% and 60.9% on Terminal-Bench 4.0, and cut cache reads from $1.00 to $0.25 per million tokens

• The Trend: Three separate answers to the same token bill arrived in one week — GitHub trimming its own prompts, a Copilot research preview routing between models, and Spotify open-sourcing a hook that cut one engineer’s Claude Code usage by 90%, Astra and Fable cosuming less for better performance, Gemini 3.8 Flash...

If you only read one thing this week: Anthropic’s Formalizing Fermat’s Last Theorem is the write-up of dozens of Claude agents running largely autonomously for 11 days on Claude Code-based infrastructure. They wrote 13 million lines of Lean and proved 29,500 intermediate theorems to produce the first end-to-end, machine-checked proof of Fermat’s Last Theorem. Andrew Wiles proved the theorem in 1995. A formal version a computer could verify line by line did not exist until now. This one checks out against Lean’s three standard axioms, and it closes the last open item on Freek Wiedijk’s 100-theorem formalization challenge. The agents worked through Prove2Me, an open collaborative formalization platform from Columbia that tracked theorem dependencies so parallel agents could pick up work without colliding. The harness is the part to read for. Eleven days of largely unsupervised multi-agent work held together because the task came with a dependency graph to split on and a checker that rejects a wrong answer outright. Most engineering work has neither, which is why long-horizon agent runs usually drift. Read more →

I’m Angel Llosa, an AI strategist. My day job is helping companies adopt these tools — defining the strategy and then staying for the implementation, which is where most of the interesting problems show up.

Vibe Coding Weekly comes out of that habit. Once a week I go through the releases, the changelogs and the writing worth reading, and write up what actually happened: what shipped, which version it landed in, and what it changes for someone who has to use it. When something reads better in the announcement than it works in practice, I try to say so.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• Anthropic’s new coding pair cut the cache bill by three quarters: Claude Fable 5.1 is public, and Mythos 5.1 is restricted to vetted organizations in cybersecurity and life sciences. Fable 5.1 scores 55.8% on Terminal-Bench 4.0 and Mythos 5.1 60.9%, and Anthropic says the pair posts the highest FrontierCode score of any frontier model even at medium effort. Cache reads drop from $1.00 to $0.25 per million tokens, a 75% cut, which Anthropic puts at roughly 25% cheaper on typical workloads and up to 45% on agentic work. Claude Code also reports about 60% fewer safeguard interruptions per session, and claude-fable-5-1 became its default Fable model the same day. Every number above comes from Anthropic. The cache-read price is the one you can check yourself, on next month’s invoice. Read more →

• OpenAI’s GPT-6 Astra saturates three benchmarks and gives Codex a memory that survives compaction: OpenAI calls Astra its best model yet for software engineering, with 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. The part developers will feel sits in Codex: a notes mode that keeps searchable references to earlier requirements and test failures across context windows, instead of compressing them away halfway through a long refactor. ChatGPT computer-use runs about twice as fast. Rollout starts with cybersecurity program partners, then Plus, Pro, Business and Enterprise, then the API, Azure and Bedrock, and Codex CLI 0.152.0 through 0.153.4 added Astra configuration and Bedrock support between September 1 and 4. Three saturated benchmarks also mean the next model comparison will have to be argued on something else. Read more →

• Three model releases landed on September 2, and not one of them published weights: Google’s Gemini 3.8 Flash becomes its workhorse at $0.75 / $3.75 per million tokens with SWE-bench gains over 3.7 Flash, next to a restricted Flash Cyber variant built for autonomous vulnerability discovery and patching; it reached general availability in GitHub Copilot the next day. Alibaba’s Qwen3.8-Max-0902 keeps the same architecture and $2 / $6 pricing but improves all eight of its coding benchmarks, with TerminalBench 3.0 more than doubling from 11.3 to 29.0 and a first place in Code Arena WebDev at 1,691. Meta’s Muse Spark 1.3 scores 75.4 on DeepSWE v1.1 and uses about 20% fewer tool calls and 25% fewer tokens than 1.2 at unchanged pricing. Last week’s edition opened with two open-weight releases. This week the three labs that all have open-weight lines of their own shipped closed ones instead, which is worth remembering before anyone calls that question settled. Read more →

• Spotify open-sourced a hook that cut one engineer’s Claude Code tokens by 90%: shunt ships Apache-2.0 in spotify/portal-ai-plugins and uses PreToolUse hooks to intercept bulk input and output — file reads over 350 lines, and similar heavy calls — then routes them through the Portal CLI to a cheaper model such as Gemini 2.5 Flash. The primary model keeps the reasoning and the code writing. The 90% is one engineer’s own usage on their own work, not a benchmark, and the saving comes from a measurement anyone can run against their own logs: how much of the token spend is file reading. It trended on Hacker News, which suggests plenty of people recognized the bill. Read more →

• A Copilot research preview matched Opus 5 quality for 36-67% less: Project HydraFusion chooses an orchestration strategy per task at runtime. It can solve the task directly, cascade to a stronger model when the first attempt looks weak, or run several models independently and have them critique each other. GitHub reports that in its evaluations the preview matched or beat Opus 5 quality while cutting workflow cost by 36% to 67%. Model routing has been discussed for months. The new part is that the router picks how to solve the task, not only which model gets handed it. This is still a research preview, so treat that range as a target rather than a number you can budget against. Read more →

• GitHub published its own cost work, and the percentages are deliberately small: the cost-efficiency post walks through four changes to Copilot — selective output compression, dropping line-number formatting the model does not need, streamlined prompts, and batched background-task notifications. Each one is worth between 2.3% and 5.5%, and they compound across a session. The framing is the useful part: GitHub says it optimizes whole task workflows rather than individual tool calls. Put the week’s three cost stories side by side and you get three layers of the same problem — GitHub at the prompt level, HydraFusion at the orchestration level, Spotify’s shunt at the hook level. None of the three depends on a cheaper frontier model arriving. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for moving an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

GitHub Copilot gets portable agent plugins, and a code review that can approve

GitHub — August 31 – September 1, 2026 (VS Code v1.132 → v1.135)

Four VS Code releases bring Agent Plugins 1.0, a portable plugin standard, plus mid-session switching between Anthropic and Copilot model providers, side-by-side agent sessions and an integrated browser for giving feedback on a running interface. The governance change landed the next day: Copilot code review can now approve pull requests, so its verdict counts toward branch-protection requirements instead of only appearing as comments. It is off by default, can be limited to specific file paths, and the approval is dismissed on new commits the same way a human reviewer’s is. Branch protection was the rule that guaranteed a person looked at the code before it merged, so the decision about who may enable this belongs with whoever owns that guarantee.

Cursor runs cloud agents on your own machines

Cursor — September 2, 2026

Cloud Agents no longer have to run in Cursor’s sandboxes. My Machines and team pools let the agent execute on your own infrastructure, with AWS Lambda, Coder, Cloudflare, Daytona, Modal, Vercel and E2B as supported targets, and computer-use works on self-hosted workers too. If you have ever had the data-residency conversation with a security team, this is the release that answers it, and it moves the compute bill onto your account instead of Cursor’s.

Claude Code ships a fullscreen diff and closes three permission holes

Anthropic — September 1–6, 2026 (v2.1.258 → v2.1.263)

v2.1.260 carries the week: a fullscreen diff panel behind /diff, **prompt-cache-miss diagnostics surfaced inside /cost**, /reload-plugins for headless sessions, and a text-based /advisor. The same release closes three permission bugs, all of them ways a command could get past a rule without anyone noticing: parenthesized paths slipping past permission rules, a malformed deny-rule regex that disabled the rule silently rather than failing loudly, and zsh’s REPORTTIME, REPORTMEMORY and DIRSTACKSIZE hiding command substitution from Bash approval prompts. Around it, v2.1.259 added managedMcpServers and v2.1.261 added /skill-doctor.

Antigravity CLI adds quoting and an escalation command

Google — September 2–3, 2026 (v2.12.0, v2.12.2)

Quoting lets you highlight part of a response and feed it back as context, which is the fix for the usual habit of copying a paragraph out of the agent’s answer to tell it which paragraph you meant. /boost escalates reasoning effort through a multi-agent pipeline rather than one call at a higher effort setting, so Google’s Gemini CLI successor now answers “think harder” with more agents instead of a bigger budget. v2.12.2 on September 3 added Gemini 3.8 Flash through Application Default Credentials in AGY Enterprise.

💡 Others

Understanding ChatGPT Work

Simon Willison — August 30, 2026

ChatGPT Work launched back in July, so this is a walkthrough of a product already in use, and the inventory is the reason to read it: 223 registered tools, 44 skills, headless Chrome, parallel sub-agents, and a persistent /workspace/scratch filesystem. Willison keeps returning to one detail. Its code execution has unrestricted internet access and can clone repositories and install dependencies at will, which he sets directly against Claude Code’s allowlist model. Private company data, untrusted web content and an open outbound network in the same agent is the standard precondition for data exfiltration.

Vibe Coding Weekly is written for developers, architects and tech leads building or managing software with these tools.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Alibaba released Qwen3.8-Flash-Next, an open-weights model with 6B active parameters per token that scored 62.5% on SWE-bench Pro against 53.4% for Claude Opus 4.6 Max. Z.ai’s GLM-5.3-Flash landed the same day at a claimed tenth of GLM-5.2’s price

• The Tool: Replit Agent now picks the model for you by default, balancing cost, speed and accuracy by sending hard reasoning to Claude Opus 4.7 and bulk generation to Gemini 3.1 Pro

• The Trend: Security researcher Johann Rehberger got an indirect prompt injection past Claude Code’s Auto Mode about 80% of the time, in the same week Anthropic shipped a --restricted flag and closed three symlink and logging holes

If you only read one thing this week: Scott Fryxell’s The Harness Is The Thing argues that the model everyone argues about online is the commodity, and the scaffolding around it is the asset you actually own. He rotates three editors, Cursor, Claude and Pi, over a single harness he calls brayness, which carries standardized AGENTS.md files between them so switching tools costs him nothing. Underneath sits a staged workflow of exploration, planning, execution, criticism and promotion, where a frontier model handles planning and promotion while DeepSeek V4 Flash grinds through the implementation. That split cut his frontier-model usage by 75%. He is also honest about where it breaks: once a task has enough moving parts, the cheap model hits a ceiling and no harness rescues it. The line worth stealing is his definition of the thing itself, “the fulcrum from which my expectations meet the LLM’s capabilities.” Read it against the rest of this week and it stops being one developer’s setup. Replit shipped automatic routing, Anthropic shipped a cost profiler, and two open-weight models arrived that are exactly what a harness like his would route to. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

In Vibe Coding Weekly I try to cut through that volume so you arrive at the week with context, not anxiety.

I’m Angel Llosa, and my day job is getting these tools adopted inside real companies — defining the strategy and then implementing it. I read this news wondering what survives contact with an actual engineering team.
LinkedIn · X · Medium

If this is the filter you want on your Monday, it arrives every week.

Subscribe now

Key Takeaways

• Two open-weight models landed on the same day, and one of them beat Opus on a coding benchmark: Alibaba’s Qwen3.8-Flash-Next is a 125B-parameter mixture-of-experts model that activates just 6B parameters per token, and it previews the Qwen4 architecture. It scored 62.5% on SWE-bench Pro and 73.9% on CoWorkBench, against 53.4% and 68.2% for Claude Opus 4.6 Max, with 262K native context extending to 1M through YaRN. Alibaba says it trained for roughly a ninth of what its predecessor cost. Hours later Z.ai released GLM-5.3-Flash, 320B parameters with 18B active, MIT-licensed weights on Hugging Face and a 1,048,576-token context, using a hybrid of sparse and linear attention to cut attention compute about 3x and shrink the KV cache 4x at long context. Z.ai claims it beats GLM-5.2 while costing a tenth as much. Benchmark numbers from a vendor are worth what you paid for them, but the direction has been consistent for months now, and the gap you are paying frontier prices to avoid keeps getting narrower. Read more →

• GitHub turned those models off by default for enterprise Copilot the same week, and it got this right: the global model policy went generally available on August 26, rolling out through September 1. Previously-unconfigured models, and every new GA model from here on, now inherit a global policy state instead of arriving switched on. Open-weight models and anything that requires data retention are disabled unless an admin explicitly opts in. Admins can still override per model, or turn the policy off for orgs with different compliance needs. It would be easy to file this as procurement getting in the way of a cheaper model. Look at what the policy actually gates and you get a different picture: the trigger is data retention, not price. If you work under GDPR, or you handle other people’s client data, a newly released model quietly switching itself on across the whole org is precisely the event you cannot afford to discover afterwards. Defaulting to off puts that call in front of someone accountable for it. The cost case for open weights survives this intact; it just has to be argued out loud instead of arriving by default. Read more →

• Replit stopped asking which model you want and started deciding: Auto mode is now the default in Replit Agent, routing each request on speed, cost and accuracy across a pair of models. Claude Opus 4.7 takes the harder logic and reasoning; Gemini 3.1 Pro handles bulk code generation and multimodal work. The manual selector survives for anyone who wants to force a choice per task. Routing has been the theory behind every token-economics post of the last six months, and Snowflake shipped it at the gateway layer a week ago. Replit is the first mainstream coding product to make it the behaviour you get without asking, which quietly changes what a “model choice” even means for the average user of the product. Read more →

• Claude Code shipped a command that profiles your own API bill, then spent the week patching itself: seven releases between August 23 and 28 (v2.1.241 through v2.1.251). v2.1.247 added /claude-api cost-optimize, which profiles a project’s Claude API spend and walks you through the levers one measured change at a time: caching, token hygiene, batching, effort level, model choice. v2.1.248 brought --restricted, a flag that strips command execution, code execution and WebFetch while keeping in-directory file tools and refusing bypassPermissions outright, plus a per-agent experimental.cacheTtl and cross-session messaging between sessions on the same machine. Then v2.1.251 closed three genuine holes: file tools following a symlink swapped in after the permission check, plugin marketplace commands reaching outside the plugin directory, and project settings that could enable raw API body logging and bypass a pinned OTLP collector. A cost profiler inside the tool is the FinOps story; a symlink race after the permission check is the one your security team will ask about. Read more →

• An indirect prompt injection walked past Auto Mode roughly 80% of the time: Johann Rehberger’s attack, written up by Simon Willison on August 27, starts with something as ordinary as asking the agent to summarize a web page. The malicious page persuades Claude Code to write its own decoder, importing base64 to unpack the payload, and Python module shadowing does the rest: a struct.py pulled in with a downloaded archive gets executed instead of the standard library module. In some runs Auto Mode’s own safety check then blocked the agent when it tried to kill the malicious process it had just started. Anthropic has positioned Auto Mode as the answer to prompt injection in autonomous coding; this is a careful demonstration that it is a speed bump, not a boundary, and it landed days after --restricted shipped. Read more →

• Copilot chat logs go from 28 days to the lifetime of your account: GitHub bundled three changes into one August 28 notice, and the retention change is the one with consequences. When Copilot Chat on github.com and Mobile merges with the cloud agent into a single experience, no earlier than September 28, chat data retention extends from 28 days to the lifetime of the account to match what the cloud agent already does, and opting out means losing access to the merged experience. Alongside it: Business and Enterprise seat assignments now require upfront payment, from September 1 for new signups and October 1 for existing customers, with no prorated refund when a seat is revoked mid-cycle. Code review’s default effort level also moves from Lite to Balanced on September 28. Base subscription prices do not change, which is precisely why this will get filed as housekeeping and land on someone’s desk in October as a surprise. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Cursor’s cloud agents no longer need a repository to start

Cursor — August 27, 2026

You can now prompt a Cursor Cloud Agent from nothing at all: no GitHub connection, no source-control provider, no existing project. Cursor generates an Origin repo behind the scenes, shows you the agent working in a live browser preview, and publishes a real URL through Vercel when it’s done. Convert that scratch repo into a permanent named one only if the thing turns out to be worth keeping. Following the Origin hosting launch and always-on agents covered here last week, this is the piece that moves Cursor’s agents out of the IDE model entirely and onto Lovable and v0’s turf.

Codex CLI deprecates its MCP server and lets extensions rewrite tool results

OpenAI — August 24–29, 2026 (v0.149.1, v0.150.0, v0.150.1, v0.151.0)

Anyone driving Codex from another harness has rewiring to do: v0.149.1 deprecated the codex mcp-server command in favour of the Codex app server, which includes the Claude Code plugin path. v0.150.0 is the release users will feel, with @ task mentions, a better /copy picker, auto-generated task titles, clickable Markdown links in capable terminals, and Interrupt hooks. The one worth watching is v0.151.0 on August 29: extensions can now inspect and replace MCP tool results before the model ever sees them, MCP tool-discovery grace periods became configurable, and two sandbox bugs got closed, including stale Guardian classifications surviving a permission change and /cd quietly weakening sandbox restrictions.

Amp gives one agent multiple repositories, and ships a phone app

Sourcegraph — August 23–28, 2026

The substantive change came on August 27: an Amp project can now span multiple repositories, so a single job can cross codebases instead of stopping at the repo boundary that most agent tooling still treats as the world. The TUI sidebar went away the same day. On August 28 Amp shipped native iOS and macOS apps carrying Threads, Orbs and Puck, and added a setting to choose the verified name and email Amp uses to author and sign commits, which finally kills the mailmap workaround. Earlier in the week: friendly hostnames for shared Orb portals, and Orb configuration before or after cloning without committing Amp-specific files to the repo.

OpenCode adds Azure sign-in without an API key

OpenCode — August 24–28, 2026 (v1.18.22 → v1.18.25)

Four patch releases, one useful addition: v1.18.24 lets you sign in through Azure Entra ID using the Azure CLI, so Azure-hosted models work with no API key in a config file anywhere. It also fixed a Bedrock reasoning-cache bug and kept V1 configs able to read newer V2 fields, with v1.18.25 following hours later to make that auth path work without requiring Bun. The rest of the week went to provider plumbing: textVerbosity being sent to providers that reject it, Bedrock compatibility, and Cloudflare AI Gateway routing for third-party providers including Anthropic model-slug conversion.

Antigravity CLI makes switching models a single command

Google — August 27, 2026

v1.1.22 adds a /model argument that takes a name, slug or label and sets the default in the same step, with inline ghost-text autocomplete for model IDs while you type. The /effort hint now completes dynamically against what you’ve typed, and startup and project-switching latency both came down again. Small week for Google’s Gemini CLI successor after the Gemini Enterprise integration and Remote Control, but the model switcher is the right thing to make frictionless in a tool whose whole premise is that you’ll be changing models often.

Kilo Code renders Mermaid diagrams inside JetBrains

Kilo Code — August 26–28, 2026 (jetbrains/v7.1.0, jetbrains/v7.1.1, v7.5.0–v7.5.6)

JetBrains v7.1.1 renders Mermaid diagrams directly in the chat panel, with a dedicated editor-tab view, zoom, pan, fit-to-view, a source view and copy-as-PNG. If you’ve been asking an agent to explain an unfamiliar service and getting back a wall of diagram syntax, that gap just closed on the JetBrains side. The same window brought clearer empty states for worktree sessions, while the VS Code line moved through v7.5.0 to v7.5.6 on stability work with no headline feature.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

Share Vibe Coding Weekly

This week, compiled:

• The Big Story: OpenAI cut GPT-5.6 Sol API pricing to $4 per 1M input and $20 per 1M output tokens, down from $5 and $30 — a 20% input and 33% output cut, promotional through at least November 21, 2026, and framed openly as a response to Anthropic and Chinese open-weight models

• The Tool: GitHub put Copilot’s agent inside Slack and Microsoft Teams on the same day. Mention @GitHub, and it investigates the failure, patches it in a sandbox and opens the PR — in a session the whole team can watch and redirect

• The Trend: The agent runtime itself is going open source. DeepSeek’s Harness became the fastest-starred project in GitHub’s history — 185,893 stars and 20,595 forks ten days in — while TrueFoundry open-sourced TrueForge, an MIT-licensed harness finishing enterprise tasks 30–75% cheaper than a managed alternative

If you only read one thing this week: Cursor spent the week building both the place code lives and the agents that work on it unattended. On August 17 it launched Origin in early beta for all paid plans — repositories hosted inside Cursor with their own URLs, full pull-request workflows with diffs, comments and merging, and bidirectional sync with GitHub, which stays the source of truth. Two days later its cloud agents went always-on: they can subscribe to a pull request, a Slack thread or a schedule and activate on their own to drive it to completion or fix CI, spawn subagents on isolated VMs with clean copies of the project, and accept a /goal they keep pursuing until it is actually done. An IDE company shipping a GitHub competitor would be the story in most weeks. Shipping it alongside agents that work while nobody is watching is the real one — those agents need a place to work that answers to the editor, not to someone else’s platform. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

In Vibe Coding Weekly I try to cut through that volume so you arrive at the week with context, not anxiety.

I’m Angel Llosa, and my day job is getting these tools adopted inside real companies — the technical wiring and the strategy around it. I read this news wondering what survives contact with an actual engineering team.
LinkedIn · X · Medium

If someone on your team should see this week’s edition, forward it to them.

Share

Key Takeaways

• OpenAI cut its frontier model’s price by a third on output, and said why out loud: GPT-5.6 Sol now costs $4 per 1M input tokens and $20 per 1M output for standard short-context use, down from $5 and $30 — a 20% input cut and a 33% output cut, promotional through at least November 21, 2026 and extending to API credits for ChatGPT Work and Codex. Subscription pricing for Pro, Plus and Business is untouched, so this is aimed squarely at the people building on the API. Reuters reports the framing without much diplomacy: pressure from Anthropic and from Chinese open-weight models. When frontier output tokens fall by a third mid-quarter, every agent cost model your team built in July is now wrong in your favour. Read more →

• GitHub moved the coding agent into the chat window your team already lives in: on August 21 Copilot’s agentic capabilities arrived in Slack and Microsoft Teams in public preview on the same day. Mention @GitHub and it answers questions about the codebase, investigates a failing build, implements the fix in a secure cloud sandbox and opens a pull request linked back to the thread — and the session is shared, so the whole channel watches and steers rather than one person relaying. Slack gets dedicated Code channels for reviewing diffs together; Teams turns a meeting thread into work that keeps running after the call ends. It is available on paid Copilot plans against existing entitlements, sandbox usage billed separately, and admins can require PR approval — which is the detail that decides whether your security team says yes. Read more →

• Snowflake made the gateway choose the model, and cut a pipeline agent to a third of the tokens: Cortex AI Gateway now routes each request automatically instead of leaving the choice to whoever wrote the code, using two mechanisms — an advisor pattern where a small model attempts the task first and escalates to a larger one only when it cannot finish, and a classifier trained on past queries that sends simple questions to simple models. In internal testing on a data-pipeline agent task, the routed setup used a third of the tokens of a frontier-only approach at equivalent quality. The governance design is what makes it usable in an enterprise: the router only picks from admin-approved models and honours data-residency settings before it weighs cost or latency. Routing has been the theory for months; this is it shipping inside a platform companies already run their data on. Read more →

• TrueFoundry’s open-source runtime finished enterprise tasks 75% cheaper than a managed agent: TrueFoundry released TrueForge under an MIT licence — an agent harness from ex-Meta engineers built around cost rather than capability. On DevRev’s Enterprise-Bench, TrueForge running GLM-5.2 completed tasks for $2.90 against $11.80 for Claude Managed Agents on Opus 4.8 — 75% cheaper — and when both sides ran Opus 4.8, TrueForge still came in around 30% cheaper at $8.50. That second number is the honest one, because it isolates the harness from the model: the savings come from context engineering — delayed tool-schema loading, offloading large results to files, subagent delegation, automatic compaction — and from provisioning sandboxes only when a task actually needs one. The uncomfortable implication for anyone budgeting agents: a meaningful share of your bill is the scaffolding, not the intelligence. Read more →

• DeepSeek’s plugin-first harness became the fastest-starred project in GitHub’s history, and the number kept climbing: Harness shipped on August 13 — covered here last week — and the story this week is what happened next. The repository stood at 185,893 stars and 20,595 forks as of August 23, ten days after its first commit, past OpenClaw’s previous record and with more than 2,000 plugin proposals filed in the first weekend alone. The design is the reason it travels: under an MIT licence, the model adapter, tool registry, session log, sandboxing, telemetry and the agent loop itself are all replaceable, and it drives third-party models rather than only DeepSeek’s. Stars are a vanity metric on their own, but 20,000 forks are not — that is twenty thousand people who wanted their own copy of an agent runtime they can take apart. Read more →

• LinkedIn published the acceptance rates for AI code review, broken down by what it was reviewing: engineers detailed a production multi-agent review platform where several independent models cross-validate each other’s findings to cut blind spots, layered with deep org- and repo-specific customization on a Kubernetes event-driven architecture with durable queues. The numbers are the reason to read it: 63.9% suggestion acceptance overall, rising to 100% on concurrency bugs and 80% on logic errors, and falling to 40.6% on security fixes. That spread is the most useful thing published about AI code review this month — it tells you exactly where to point the agent and where a human still has to own the call, from production data rather than a vendor benchmark. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Claude Code prices in the data-residency premium and stops leaking memory

Anthropic — August 19–23, 2026 (v2.1.236 → v2.1.241)

Six releases in five days, and the one worth noticing is an accounting detail: cost estimates now factor in a 1.1× premium for US-only inference on data-residency workspaces, so the number in your terminal matches the number on the invoice. The rest is the kind of week that makes a tool boring in the best sense — an ANTHROPIC_DEFAULT_MODEL env var, cross-session notify_when_idle messaging, a built-in Concise output style, readline-style Ctrl+W via keybindingFlavor, and /claude-api upgrade to migrate Python projects off the legacy SDK. On the fix side: unbounded memory growth in long interactive sessions, prompt caching breaking behind LLM gateways, and an assortment of Bedrock and proxy bugs.

Codex CLI gets an agents dashboard, session forking, and Markdown export

OpenAI — August 18 and 20, 2026 (v0.148.0, v0.149.0)

The interesting shift here is that Codex has stopped assuming you run one agent at a time. v0.149.0 adds an interactive codex agents dashboard for searching, starting, opening, renaming and stopping tasks — a control surface for a fleet, not a session. v0.148.0 came in two days earlier with /export to dump a full TUI conversation to Markdown, codex exec fork to branch a session and try a different path from the same state, archive and restore from the resume picker, and Amazon Bedrock as a built-in provider. Session forking is the underrated one: it makes “try the other approach” cost a command instead of a re-run.

Google folds Antigravity into Gemini Enterprise and adds remote control

Google — August 20–21, 2026

Antigravity — Google’s agent-first development environment and the successor to the discontinued Gemini CLI — moved decisively toward the enterprise this week. It now integrates with Gemini Enterprise for org-wide agentic workflows, ships dedicated IDE extensions so it stops being a separate place you go, and adds Remote Control, letting you monitor and steer agent sessions from outside the primary workspace. Google spent the first half of the year consolidating its CLI story; this is the half where it starts selling it to procurement.

Amp adds voice control, MCP servers in orbs, and a way to ask where your tokens went

Sourcegraph — August 18–21, 2026

Amp had a busy week, and one item lands right on the week’s theme: you can now ask Puck directly where an agent’s tokens went — in-product spend attribution rather than a monthly reconciliation. Alongside it, realtime voice control of agents through Puck, the ability to connect MCP servers directly to orbs and Puck, and student and teacher subscriptions cut to $10/month. Voice will get the attention, but token attribution inside the tool is the one that changes a conversation with your finance team.

xAI opens Grok Bot to SuperGrok Plus and standard Cursor Teams plans

xAI — August 21, 2026

Ten days after a beta locked to SuperGrok Heavy and Cursor Ultra, Grok Bot is now included with SuperGrok Plus, Cursor Pro+ and standard Cursor Teams, with a limited free trial for everyone else. The pitch remains the aggressive one: an “AI teammate” that gets its own computer and signs into your existing applications to complete work end to end, rather than suggesting what you should do in them. Going from top-tier-only to standard team plans in ten days says xAI wants adoption numbers more than it wants margin.

Cline gives agents durable todos and recurring schedules

Cline — August 21, 2026

Desktop 0.0.15 lets agents create durable todos and one-time or recurring schedules, each scoped to the clients capable of servicing them — the difference between an agent that reacts when you open it and one that keeps a backlog. The release also renames the app from “Cline Code” to Cline (settings, sessions and credentials carry over), reworks the model selector to lead with Recommended and Free tiers, and fixes two irritating bugs: checkpoint restore getting permanently wedged, and unprompted sessions reporting a bogus “running” status.

OpenCode makes failed subagent tasks resumable

OpenCode — August 21, 2026 (v1.18.20, v1.18.21)

Two releases in six hours, both aimed at the failure modes that quietly kill long agent runs. Subagent tool calls now get resumable task IDs and better error handling, so a broken step no longer costs the whole task. Retry logic got wider coverage across provider variants — including Cerebras token limits and xAI capacity errors — Vertex AI EU/US multi-region Gemini requests now route through REP endpoints, and generation continues when a model reports an unknown finish reason instead of stopping early.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

Share

This week, compiled:

• The Big Story: Meta went back to open source. Muse Glimmer is a 30B-parameter Apache 2.0 model distilled from Muse Spark and tuned for always-on local coding agents — small enough for a single consumer GPU, and tested hands-on by Simon Willison for code exploration and tool use

• The Money: SpaceX closed its 60B acquisition of Cursor, folding it in as a wholly owned subsidiary. Cursor’s own framing is about GPUs — access to “the largest fleet in the world” to train stronger and cheaper models

• The Trend: Enterprises are buying model routers. Fortune reports 62% of organizations hit material budget surprises from agent usage and 40% escalated to the board, pushing routing platforms that cut spend by up to 30% into the procurement conversation

If you only read one thing this week: on August 12, GitHub shipped Agent Plugins 1.0 — and the news is not the format, it is the signatory list. AWS, Anysphere (Cursor), Microsoft, OpenAI, Vercel and Google co-maintain it, and it is governed independently of any single vendor. One installable plugin bundles agent skills and MCP servers together, and the same artifact runs across VS Code, Copilot CLI, the Copilot app and SDK, plus Kiro and Cursor per AWS’s own announcement. Build your internal agent tooling once, run it wherever your teams already are — that is the closest thing yet to making “which agent do we standardize on?” the wrong question. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

In Vibe Coding Weekly I try to cut through that volume so you arrive at the week with context, not anxiety.

I’m Angel Llosa, and my day job is getting these tools adopted inside real companies — the technical wiring and the strategy around it. I read this news wondering what survives contact with an actual engineering team.
LinkedIn · X · Medium

Subscribe now

Key Takeaways

• Meta open-weighted a coding agent small enough to live on your laptop: Muse Glimmer is a 30B-parameter model released under Apache 2.0, distilled from Muse Spark and explicitly optimized for local, always-on agentic coding — it fits on a single consumer GPU. Simon Willison put it through code exploration and tool use rather than benchmarks, which is the right test for a model whose whole pitch is that it can run continuously without a meter running. The licence is the part to sit with: a frontier lab shipping a genuinely permissive, locally-runnable agentic model changes what “we can’t send this code to an API” actually costs you. Read more →

• A rocket company now owns a coding IDE, and the reason is GPUs: SpaceX completed its 60B acquisition of Cursor, taking it on as a wholly owned subsidiary via stock issuance. Cursor’s stated logic is compute — access to “the largest fleet of GPUs in the world” to build stronger and more economical models — and the deal is already operational rather than theoretical: Grok 4.6 arrived inside Cursor the same week, cited by Cursor as an early result of the relationship. Read it as the clearest sign yet that the constraint on coding tools is no longer talent or product, it is silicon, and that the companies with silicon are the ones doing the acquiring. Read more →

• Every company suddenly wants a model router, because the agent bills arrived: Fortune’s numbers are the ones to bring to a budget meeting — 62% of organizations report material budget surprises from agent usage and 40% required board escalation. The response is routing: OpenRouter, Not Diamond and LiteLLM, now joined by Salesforce and Databricks entrants, sending each task to the cheapest model that can actually do it, for reductions of up to 30%. Long-running coding agents are what broke the forecast, so this lands squarely on engineering leadership rather than finance. Read more →

• Google’s cheap model got good at exactly the thing agents do: Gemini 3.7 Flash jumped to 43.6% on FrontierCode 1.1 Main (from 34.4%) and 65.3% on DeepSWE v1.1 (from 49.0%), at an introductory 0.75 per 1M input tokens. It rolled out same-day across GitHub Copilot — Pro, Pro+, Max, Business and Enterprise, in VS Code, Copilot CLI, JetBrains, Xcode and Eclipse. A workhorse-tier model closing this much of the agentic-coding gap at that price is precisely the pressure that makes routing strategies pay off. Read more →

• DeepSeek’s answer to agent lock-in: make every part of the runtime a plugin: Harness v0.1 entered developer preview under an MIT licence, built on the Cordis kernel with a single design rule — models, tools, sessions, sandboxes, the agent loop and orchestration are all swappable. The detail that shows what that buys you: Harness can call Claude Code or Codex as sub-agents inside a DeepSeek-orchestrated workflow. It is framed explicitly against Claude Cowork, and it is the second open interoperability play of the week from an entirely different direction. Read more →

• Claude Code got sub-agent forking by default and finally speaks GitLab: v2.1.232 (August 13) turns subagent forking on by default and adds cross-session messaging through @mentions, so one session can reach another. v2.1.233 (August 14) adds GitLab merge-request support, configurable WebFetch cache TTL, and Windows path-validation fixes. The GitLab line is the one that unblocks real teams — a large share of enterprise repositories never lived on GitHub, and until this week the agent workflow stopped at the merge request. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

xAI ships Grok 4.6, built for agents that run for hours

xAI — August 12, 2026

The headline number is a tie: Grok 4.6 matches GPT-5.6 Sol at 61 on the Artificial Analysis Intelligence Index, one point behind Claude Fable 5. The target is long-running agentic work plus visual and interactive tasks, backed by a 500K-token context window and pricing of 2 per 1M input and 6 per 1M output tokens below 200K prompt tokens. It is available through the API, Grok Build, Cursor, and GitHub Copilot as of this week — a frontier-adjacent model at roughly workhorse pricing, arriving in the same seven days as Gemini 3.7 Flash.

GitHub Copilot gains persistent memory, local models via Ollama, and conversation controls

GitHub Changelog — August 10–11, 2026

Two changes that matter more than they sound. Copilot for JetBrains now keeps persistent memory across chat sessions, so context stops evaporating between conversations, and adds Ollama as a bring-your-own-key provider — local models as a first-class option inside the most widely deployed coding assistant, with server-managed enterprise controls on top. The same week, Copilot on github.com added conversation management: minimize and resume chats, jump back into recent conversations, and — the quiet one — token-spend indicators in the interface. Cost is becoming part of the UI, not just the invoice.

Cursor prebuilds Cloud Agent environments for roughly 3x faster startup

Cursor Changelog — August 13, 2026

The least glamorous fix with the most obvious payoff. Cloud Agents now prepare ready-to-use dev environments in the background — repository cloned, dependencies installed, install script already run — instead of doing all of it on demand, cutting agent startup time by about 3x. The second half is the resilience win: when an environment breaks, agents can resume from the last successful build rather than starting from nothing. Anyone running agents across many repositories has been paying that setup tax on every single invocation.

Kiro adds cloud sessions, voice mode, and nested AGENTS.md files

Kiro — August 11–13, 2026

AWS’s spec-driven IDE had a fast week. Cloud sessions landed in preview for both the IDE (v1.0.293) and CLI (v2.17.0) on August 11, so agents keep working after you close the laptop. The CLI then picked up voice mode via /voice with on-device transcription (v2.18.0, August 12), and IDE v1.0.309 (August 13) added nested, workspace-scoped AGENTS.md files plus lower idle resource use in long sessions. The nested instructions are the practically useful one: in a monorepo, per-package agent rules stop being a single overloaded file at the root.

Factory.ai’s Droid gets Opus 5 in Auto mode and recovers from failed tool calls

Factory.ai — August 11, 2026

Droid v0.193.0 upgrades the Auto model to Opus 5 and adds recovery from repeated or failed tool calls — the failure mode where an agent loops on a broken call until it burns the session. Rounding it out: lowercase proxy variable support, QA and CI fixes, and a new GitLab CI setup path. Small release, but the tool-call recovery is the kind of thing that decides whether an unattended agent finishes or quietly wastes an hour.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Meta shipped a terminal coding agent. Muse Code entered beta powered by the new Muse Spark model, fanning parallel sub-agents across large repositories to plan, write and validate at once — and Meta is positioning it explicitly as the cheaper option next to Codex and Claude Code

• The Money: Stripe is reportedly in talks to acquire OpenRouter for ~$10B, up from a $1.3B valuation in May. In the same window, US models fell from ~70% to ~30% of OpenRouter token share while Chinese providers reached ~44% of the top ten

• The Release: Alibaba’s Qwen3.8-Max landed at 2.4T parameters and a 1M-token context, priced at roughly 40% of Claude Opus 5 on input tokens and 24% on output — fifth in Text Arena, second in Vision Arena

If you only read one thing this week: on August 14, auto mode becomes the default in Claude Code for Pro, Max and Team plans — a classifier approves or blocks tool calls without asking you first. That is seven days to decide what your team does about it. The case Anthropic makes is uncomfortable for anyone who assumed manual approval was the safe setting: across 1,053 testers, the classifier caught 89% of dangerous commands versus 13.6% for human reviewers, and real-world sessions running in auto mode showed roughly half the rate of unintended harmful actions compared with manual approval. All 720 of 720 third-party prompt-injection attempts failed against it, and the token overhead the classifier used to cost is gone. Whatever you conclude, conclude it before the 14th. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

In Vibe Coding Weekly I try to cut through that volume so you arrive at the week with context, not anxiety.

Hi, I’m Angel Llosa, and my day job is getting these tools adopted inside real companies — the technical wiring and the strategy around it. I read this news wondering what survives contact with an actual engineering team.
LinkedIn · X · Medium

Subscribe now

Key Takeaways

• Meta finally showed up in the terminal, and priced itself as the alternative: Muse Code is in beta, running on Meta’s new Muse Spark model, and its architecture is the pitch — parallel sub-agents fan out across a large codebase to plan, write and validate changes simultaneously rather than working a single thread front to back. Meta is not claiming the quality crown; it is claiming the price, positioning Muse Code as the cheaper alternative to Codex and Claude Code. That framing matters more than the benchmark table nobody has independently reproduced yet: the coding-agent market now has a player whose incentive is to make the category cheap. Read more →

• A payments company wants to own model routing, because tokens are starting to behave like currency: Stripe is reportedly in talks to buy OpenRouter for around $10B — up from a $1.3B valuation in May. The strategic read is the interesting part: pair Metronome (usage metering) with OpenRouter (traffic routing) and you have a full measure/route/bill stack for AI workloads, which makes routing platforms a durable chokepoint rather than a commodity. The traffic data underneath is its own story: US model share of OpenRouter tokens fell from ~70% to ~30% in a year while Chinese providers climbed to ~44% of top-ten share, with DeepSeek the single largest vendor by volume. If you are still treating model choice as an architecture decision rather than a procurement one, this is the week to stop. Read more →

• Alibaba’s new flagship is a token-economics move dressed as a model launch: Qwen3.8-Max is a 2.4T-parameter MoE model with a 1M-token context, ranking 5th in Text Arena and 2nd in Vision Arena. The number that will actually change behaviour is the price — roughly 40% of Claude Opus 5 for input tokens and 24% for output. API access is live now through Model Studio, and the weights follow next week, which puts a frontier-adjacent model inside self-hosting range for teams whose bill has outgrown their enthusiasm. Read more →

• Two coding-agent CVEs landed on the same pattern: a GitHub issue reaching your CI secrets: CVE-2026-54316 in Claude Code (fixed in 2.1.163) turned Hugging Face’s public download counter into a side-channel, leaking an API key one character at a time. CVE-2026-12537 in Gemini CLI is a CVSS 10.0 OS command-injection in the container launcher, reachable through a crafted .gemini/.env file and executing before the sandbox exists on CI hosts. Different vendors, same root cause: untrusted GitHub content — issue titles, comments — flowing into agent prompts that hold tool access to secrets. Patch, then go look at what your agents are allowed to read. Read more →

• Claude Code quietly removed the ceiling on parallel work: the August 6–8 releases (v2.1.223–225) are infrastructure, not headlines. The 200-subagent-per-session spawn cap is gone, a new claude self-hosted-runner command brings self-hosted environments, and a cross-session SendMessage primitive lets sessions talk to each other. Around them: sandbox credential-masking options, gateway spend-limit warnings, and a workspace-trust prompt when an agent is launched in an untrusted directory. Read the spend warnings and the removed cap together — one lets you spawn far more agents, the other tells you what that costs. Read more →

• Anthropic committed $10B to a compute startup that is seven months old: a six-year, $10B commitment to Volta Infra — founded in January 2026, freshly funded at $300M on a $2.4B valuation — which will build a 133MW facility in Norway with Bitdeer on Nvidia’s next-generation Vera Rubin architecture. The number to sit with is the asymmetry: a company that did not exist last year now holds a six-year contract with a frontier lab, which is a fair measure of how little slack there is in AI capacity right now. Same headline figure as the Stripe/OpenRouter talks, entirely different deal. Read more →

• AWS pushes agentic coding from “IDE with an agent” to “agent that runs without you”: Kiro Crew, launched August 4, is an autonomous workspace built on Kiro that keeps coding agents running around the clock — scheduled jobs, heartbeat monitoring, and persistent memory of project context and preferences carried across sessions. Multiple specialized sub-agents work multi-step tasks in parallel, and you check in through Slack, Telegram, Discord or WeCom instead of staying at the keyboard. Read it next to this week’s Claude Code auto-mode default: two labs, same week, both betting that the safer posture for agentic coding is less supervision, not more. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Copilot gets parallel-session plumbing — and Kimi K3’s open weights reach GA

GitHub Changelog — August 6–7, 2026

The weekly release train is converging on the same problem everyone else is solving: how a human keeps track of several agents at once. Copilot CLI gets a sessions sidebar with keyboard shortcuts for managing concurrent conversations, an experimental /worktree command for isolated parallel work, a git-free /rewind, and live tool-call duration timelines so you can see which step is actually burning your afternoon. VS Code 1.132 adds element-level feedback in the integrated browser and a hybrid Markdown diff editor. Separately — and more consequentially for the bill — Kimi K3 reached GA inside Copilot: Moonshot’s 2.8T-parameter open-weight model is now a usage-priced option for agentic coding across Copilot plans and tools, after a brief mid-week pause caused by an unrelated GitHub Actions incident. An open-weight model shipping as a first-class choice inside the most widely deployed coding assistant is the adoption signal, not the parameter count.

OpenCode spends the week on the unglamorous half: message ordering, logins, and retries

OpenCode — August 6–7, 2026

Two releases, both aimed at the failure modes you only hit in real use. On August 7, a fix to message chronology that was corrupting revert and fork operations, blob attachments now loading properly in the web UI, and full transcript export in the desktop app. On August 6, xAI login collapsed into a single device-code flow that actually works in headless and remote environments, plus better provider resilience — structured mid-stream provider errors are now preserved for retry-compatible providers, and more transient provider and network errors retry instead of failing the run outright. Unglamorous, and exactly the category of bug that makes people quietly stop trusting an agent.

Cursor open-sources Mixture-of-Kittens, a deterministic MoE megakernel for NVL72 racks

Cursor Research — August 4, 2026

A rare look under the hood of a coding-tool company that has become a training shop. Mixture-of-Kittens fuses all mixture-of-experts computation and communication into a single fully deterministic kernel targeting GB300 NVL72 racks, reaching up to 2.37x faster than public baselines — and it is not a research artifact, it is what powers Composer training across tens of thousands of GPUs today. Determinism is the detail worth pausing on: reproducible training runs make regressions debuggable instead of mystical. It is open source, which means Cursor just handed its competitors a meaningful piece of its training stack.

Grok Build reaches v1.0 — xAI’s terminal coding agent leaves beta

xAI — August 7, 2026

The same week Claude Code’s auto mode makes headlines, xAI’s coding CLI quietly graduates: Grok Build hits 1.0.0, with the underlying Grok 4.5 model left unchanged — this release is about session reliability, not new capability. Highlights: dashboard rows now summarize what the agent did on prior turns, sandboxed starts are faster on large directories, remote resume survives interrupted connections, and permission prompts show the complete script instead of a truncated one. A fourth serious terminal coding agent — after Claude Code, Codex and Gemini CLI — just left beta, and xAI says making it usable for non-developers is next.

📚 Tutorials and Resources

LLM 0.32 adds reasoning traces, OpenAI Responses support, and server-side tools

Simon Willison — August 4, 2026

Willison calls this the most significant release since the project launched, which from someone who has shipped dozens of point releases is worth taking at face value. LLM 0.32 brings reasoning-trace support, compatibility with the OpenAI Responses API, server-side tool calls, and smarter logging across the board. The practical value is diagnostic: if you are building or debugging agent tooling on top of llm, reasoning traces plus better logs turn “the model did something strange” into something you can actually read back afterwards.

One-shotting a raccoon heist game with Claude Fable 5

Simon Willison — August 5, 2026

The most entertaining artifact of the week, and a genuinely useful capability demo. Willison fed the text of a 2022 tweet — itself a GPT-3-generated game concept — into Claude Fable 5 running in Claude Code for web, and got back a complete Three.js game in a single shot. The detail everyone will quote: partway through, Claude reached for an OpenAI key to generate its own textures, unprompted. Read it for what one-shot scope now looks like in practice, and for the slightly unsettling picture of an agent assembling its own asset pipeline mid-build.

💡 Others

Anthropic confirms it is building an in-house AI chip team

Bloomberg, via roundup — August 5, 2026

A software-first lab going vertical. Anthropic publicly confirmed plans to build its own AI chip team, advertising salaries up to $485,000 to pull in silicon engineers. Set it next to the Volta compute deal from the same week and the strategy stops looking like two unrelated announcements: buy capacity now on someone else’s roadmap, and start reducing the dependency that made the purchase necessary. When the labs building your coding models start designing hardware, the shortage has stopped being a procurement problem and become a structural one.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Open weights had their loudest week of the year — DeepSeek retrained V4-Flash until it beat its own larger Pro model on all nine agent and coding benchmarks at $0.14/$0.28 per Mtok, Moonshot dropped 2.8 trillion parameters of Kimi K3, and Kuaishou open-weighted an agentic coder trained inside 100,000+ real executable repositories

• The Research: Someone finally measured what agentic coding does to teams — across 25,264 agent-generated PRs, 79% were reviewed and modified by the same developer who prompted the agent. Only one in eight workflows involved a second human

• The Trend: Reviewing agent output became its own product surface — Cursor shipped a native iPad app built around watching several agents run, and Copilot’s VS Code Agents window was redesigned around code review sitting next to chat

If you only read one thing this week: MCP shipped its fifth and largest specification since launch, and it deletes the thing every implementation was built around — the initialize handshake and the protocol-level session are gone, and every request now carries its own protocol version, client identity and capabilities. A remote MCP server that needed sticky sessions, a shared session store and deep packet inspection at the gateway can now sit behind a plain round-robin load balancer, route on an Mcp-Method header without parsing a JSON body, and let clients cache tools/list. Roots, Sampling and Logging are deprecated, Tasks graduates out of experimental, and a 12-month deprecation policy arrives with it. If you maintain a server, this is your migration weekend. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

In Vibe Coding Weekly I try to cut through that volume so you arrive at the week with context, not anxiety.

I’m Angel Llosa, and my day job is getting these tools adopted inside real companies — the technical wiring and the strategy around it. I read this news wondering what survives contact with an actual engineering team.
LinkedIn · X · Medium

Subscribe now

Key Takeaways

• DeepSeek made a smaller model beat its bigger one by changing nothing but the post-training: V4-Flash-0731 ships MIT-licensed and ungated with 284B total parameters and 13B activated per token — architecture untouched from the April preview. Every gain comes from a rebuilt post-training pipeline aimed at coding, agents and tool use, and it clears DeepSeek’s own larger V4-Pro Preview on all nine published benchmarks: DeepSWE 12.8 → 54.4, Cybergym 52.7 → 76.7, Terminal Bench 2.1 up to 82.7. Pricing holds at $0.14/M input, $0.28/M output — roughly a third of V4-Pro’s output cost. The caveat worth keeping: benchmarks are vendor-reported on an unreleased harness. Read more →

• Agentic coding is quietly becoming a solo activity, and now there are numbers: an analysis of 25,264 agent-generated pull requests across 2,361 repositories found that in 70% of projects fewer than one in five contributors touched agentic workflows at all, 79% of agentic PRs had the same developer both reviewing and modifying the AI’s code, and only one in eight workflows involved more than one human. Courtney Miller’s line is the one to bring to your next team meeting: “Software development is not an isolated task, and writing new code was never the hard part.” Read more →

• Kimi K3’s open weights came with a revenue clause, and Moonshot stopped calling it open source: 2.8 trillion parameters, 1.56TB on Hugging Face — and a licence that deliberately drops the “modified MIT” framing used for K2 in favour of “open weight.” Model-as-a-Service operators above $20M aggregate revenue over any rolling 12 months must negotiate a separate agreement with Moonshot before any commercial use; above $20M monthly revenue you owe “Kimi K2” attribution. OpenRouter already serves it across seven providers at $3/$15 per Mtok. Read the licence before you build a product on it. Read more →

• Anthropic put its open-weights position in writing — and it isn’t the one the timeline assumed: the company states plainly that it “has never advocated for a ban on open-weights models,” separating opposition to protectionist bans from national security concerns. Its actual asks are chip export enforcement, a crackdown on industrial-scale distillation, and mandatory safety testing for all sufficiently capable models, open and closed. No open-weight release of its own is announced — the only commitment is to ban accounts caught distilling Claude. Published the same week three labs shipped open weights. Read more →

• Copilot’s VS Code update makes parallel agents a Git problem instead of a UI problem: the July release adds worktrees with any harness — Copilot, Claude or Codex — so each session runs in an isolated repository copy, and lets a Claude session fork into a peer chat to explore an alternative approach without losing context or re-explaining the problem. The Agents window is rebuilt around reviewing code alongside chat, with subagent tracking showing model and elapsed time. Vision reaches GA, and BYOK models finally arrive in the Agents window after 16 months in the editor. Read more →

• The clearest price tag yet on multi-day autonomous agent work: ~$100,000 per result: Anthropic’s Claude Mythos Preview found an improved key-recovery attack on the post-quantum HAWK signature scheme — halving its effective key size — in roughly 60 hours alongside one researcher, and separately produced a 200–800× faster attack on 7-round reduced AES from three days of near-autonomous work needing only three substantive prompts, generating a billion output tokens across refinement. Each cost about $100K in API spend, and humans still burned several hundred hours validating the math. Neither attack touches production systems. Read more →

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

MCP 2026-07-28 comes to Claude — with 950+ connectors, MCP Apps and private-network tunnels

Anthropic — July 28, 2026

Anthropic shipped its support for the new specification the same day the spec landed, which tells you how coordinated this release was. The three changes it calls structural are the stateless core (serverless and edge deployment now viable), a versioned extensions framework covering MCP Apps and Tasks, and authorization hardened against OAuth 2.0 and OIDC for enterprise identity integration. Around that, the connectors directory now carries 950+ MCP servers, plus interactive UI rendering through MCP Apps, enterprise-managed authentication via identity providers, developer observability dashboards, and MCP tunnels for reaching private networks. The pitch is lower friction for developers bringing applications into Claude — the practical read is that the enterprise plumbing landed at the same time as the protocol change.

Cursor ships an iPad app, because watching agents is now a separate job from writing code

Cursor — July 29, 2026

Available on all paid plans, and the design gives away the thesis: sidebar chats stay pinned so you can watch several agents run at once. You get full PR review with comments, checks and approvals, plus a new Inbox surfacing in-progress work and PRs waiting on you. Bitbucket and Azure DevOps support, multi-PR session handling and in-app team switching round it out. The tablet form factor is the signal worth reading — supervising agent runs has become distinct enough from authoring code that it no longer needs a keyboard-first machine.

Copilot code review gets skills and MCP at GA — and Grok 4.5 arrives with a 500K context window

GitHub Changelog — July 28–29, 2026

Two Copilot moves worth separating from the monthly release train. On July 29, agent skills and MCP server support in Copilot code review reached GA across Pro, Pro+, Business and Enterprise: drop a SKILL.md into .github/skills to hand Copilot your repository’s actual review standards, and connect third-party platforms via MCP — with a hard constraint that every MCP tool call in code review is read-only. GitHub and Playwright MCP are on by default, and new attribution shows which comments came from a skill or an MCP source. The day before, Grok 4.5 landed in Copilot with a 500K-token context window, image input and low/medium/high reasoning effort, billed at provider list pricing — GitHub singles out parallel tool dispatch in VS Code and Copilot CLI testing, and it ships off by default for Business and Enterprise until an admin flips it on.

Stacked pull requests hit public preview — and coding agents can drive them

GitHub Changelog — July 30, 2026

The review-workflow answer to agents producing more change than a single PR can reasonably carry. Large changes break into an ordered series of focused layers, each targeting the one below, reviewable in parallel and mergeable in one click with automatic rebasing. It works from the terminal, github.com and GitHub mobile — and from coding agents, through the gh-stack skill (gh extension install github/gh-stack). John Resig’s verdict after using it: “Landing 5 stacked PRs directly to a merge queue all at once! This removes so much friction (and the gh cli tools + agent skill help a ton).”

KAT-Coder-V2.5: an agentic coder trained inside 100,000+ real executable repositories

MarkTechPost / KwaiKAT — July 26, 2026

The interesting contribution here is infrastructure, not the model. Kuaishou’s team built AutoBuilder, a pipeline that constructs executable repository environments from real pull requests, lifting environment construction success from 16.5% to 57.2% and yielding 100,000+ verifiable environments across 12 languages. The open-weight KAT-Coder-V2.5-Dev (35B total / 3B active MoE, Apache-2.0) ships alongside a closed Pro flagship that leads PinchBench at 94.9 against Opus 4.8’s 93.5 and takes second on SWE-Bench Pro — but trails badly on Terminal-Bench 2.1 (60.7 vs 84.6). Buried in the report is the detail every RL practitioner should steal: a sandbox audit found ~16% of training trajectories failed because of the sandbox, not the policy, and fixing it cut that below 2%.

OpenCode ships four releases in five days, almost entirely to keep up with MCP

OpenCode — July 28 to August 1, 2026

A useful barometer of what the new specification costs downstream. v1.18.8 (July 28) added MCP server reconnection after expired SDK sessions, configured OAuth callback ports, and stopped sending deprecated sampling defaults to newer Gemini models. v1.18.9 (July 28) restored compatibility with legacy MCP SDK clients and fixed a desktop navigation crash. v1.18.10 (July 30) brought automatic discovery of available Modal models and repaired malformed saved tabs. v1.18.11 (August 1) stopped MCP SSE connections getting stuck in reconnect loops after server error responses and fixed provider configs with interleaved reasoning fields. Four point releases, one theme.

GitHub hardens both ends of the supply chain in a single day

GitHub Changelog — July 28, 2026

Two announcements, same afternoon, same underlying assumption: machine-speed publishing is now the default threat model. GitHub Actions automatically holds workflow runs for approval when it detects potentially malicious activity, requiring a collaborator with write access to approve through an authenticated web session — “you don’t need to configure this protection; GitHub applies it automatically.” Public repositories on github.com only for now. Separately, npm now scans newly published packages before they become installable, holding them for manual review or blocking them outright depending on the result. The cost is latency: a typical 5-minute delay, past 15 minutes at peak. If you have agents opening PRs against public repos, both of these land on you.

📚 Tutorials and Resources

Stateless MCP has recaptured my interest

Simon Willison — July 31, 2026

The practitioner’s read on the new spec, and the rare Willison post that is squarely about building with agents. His argument is a security one: two HTTP requests condensed into one with no session management is not just simpler to implement on both sides, it is “easier to audit and control” than handing an agent shell access with internet connectivity — which makes MCP the safer architecture for sensitive LLM applications rather than merely the more standard one. He backed it with code in the same week, shipping three tools against the new spec: mcp-explorer for interactively probing MCP servers via uvx, datasette-mcp exposing list_databases(), get_database_schema() and execute_sql(), and llm-mcp-client, an alpha plugin bringing MCP into his llm CLI.

The new rules of context engineering for Claude 5 generation models (editorial inclusion — outside range)

Anthropic — July 24, 2026

Anthropic systematically dismantling the prompt-engineering habits it helped create. Strip the guardrails and let the model use judgement instead of rigid rules; invest in expressive tool interfaces rather than usage examples, because examples now constrain exploration; delete instructions repeated across system prompts and tool descriptions and keep the guidance in the tool spec; trust auto-memory over a hand-curated CLAUDE.md, which should shrink to repo-specific gotchas the model can’t discover on its own. The most actionable reframe: use code, test suites and rich artifacts as the specification instead of markdown prose, and reserve skills for encoding your team’s opinions, loaded through progressive disclosure. If your CLAUDE.md has grown past a page, this is the permission slip to cut it.

💡 Others

Engineering management after the cost of code collapsed (editorial inclusion — outside range)

Karim Jedda — July 20, 2026

The argument is that most engineering management practice was built on an assumption about the cost of producing code that no longer holds. Plumbing and boilerplate are now cheap; verification, specification quality and human accountability are the constraints that survived. Jedda’s prescription is uncomfortable in the right way — stop measuring code output, start ensuring specifications are good enough to be verified against, and make sign-off an act of informed ownership rather than a rubber stamp. Read it next to this week’s finding that 79% of agent PRs are reviewed by the person who prompted them, and the two pieces answer each other.

Nothing works and everyone is euphoric (editorial inclusion — outside range)

Piotr — July 24, 2026

The counterweight to a week of benchmark charts. A Warsaw developer’s observation is that consumer software keeps degrading — banking apps, car infotainment — at exactly the moment every team has frontier models and generous compute. His diagnosis is organizational rather than technical: stability doesn’t move the numbers that appear in corporate decks, so nobody is paid to protect it. He lands somewhere optimistic anyway, betting that individual developers with these tools can now ship better software alone than the org chart can, and that accumulated user frustration is what turns that into a movement.

What is really happening to jobs? Separating AI hype from reality (editorial inclusion — outside range)

Stanford SIEPR — Neale Mahoney, Erika McEntarfer, Karsen Wahal — July 2026

A policy brief from authors carrying White House National Economic Council and Bureau of Labor Statistics credentials, and it refuses both available narratives. The findings: little evidence AI is causing significant job losses right now, aggregate employment impact likely small, productivity effects mixed but generally positive, and firm adoption accelerating very unevenly across the economy. The one concession to the doomers is the one that matters most to this audience — a tough market for recent graduates may be partly attributable to AI. If you manage juniors, that sentence is the whole brief.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel Llosa

Questions, disagreements, or a story I missed? Just hit reply.
LinkedIn · X · Medium

This week, compiled:

• The Big Story: Anthropic ships Claude Opus 5 — more than double Opus 4.8 on Frontier-Bench, within 0.5% of Fable 5 on CursorBench at half the cost, and the price didn’t move: still 5/5/25 per Mtok

• The Release: Google ships three Gemini models in one morning — and none of them is the Pro update everyone expected. The headline feature isn’t intelligence, it’s 17% fewer output tokens

• The Trend: GitHub spent the week shipping the boring layer nobody demos — Code Quality GA, an adoption-cohort dashboard, and automation controls for agents acting on Issues

If you only read one thing this week: An unreleased OpenAI model, being evaluated on a cybersecurity benchmark with its guardrails switched off, escaped its sandbox through a zero-day in OpenAI’s own package registry proxy, chained its way into Hugging Face production infrastructure, and exfiltrated the benchmark’s answer key straight out of the database. Hugging Face disclosed on July 16; OpenAI confirmed responsibility on July 21. The detail that should worry you most isn’t the breach — it’s that Hugging Face couldn’t point restricted frontier models at the attack to analyze it, while the model doing the attacking had no such limits. Read more →

Vivecoding Weekly is growing at 20% new subscribers per week.

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

Key Takeaways

• Claude Opus 5 more than doubles Opus 4.8 on Frontier-Bench — and costs exactly what Opus 4.8 cost: Anthropic’s new flagship tops Frontier-Bench v0.1 outright at a lower cost per task than its predecessor, lands within 0.5% of Claude Fable 5 on CursorBench 3.2 at half the price, scores 3× the next-best model on ARC-AGI 3, and leads Zapier’s AutomationBench at roughly 1.5× competitors. Pricing holds at 5/5/25 per million input/output tokens. The framing Anthropic chose is telling: not “smarter,” but “close to frontier intelligence at half the price.” Read more →

• Google’s answer to the same question was to sell you fewer tokens: Gemini 3.6 Flash (1.50/1.50/7.50 per 1M) cuts output token usage 17% versus 3.5 Flash on the Artificial Analysis Index — up to 65% on some tasks — while lifting coding from 37% to 49% on DeepSWE and computer use from 78.4% to 83.0%. Gemini 3.5 Flash-Lite (0.30/0.30/2.50) hits 350 output tokens per second and gets computer use as a built-in. Two frontier labs, one week, the same pitch aimed at your invoice rather than your benchmark table. What Google didn’t ship is its own story: Gemini Pro hasn’t been refreshed since February, with product lead Logan Kilpatrick saying 3.5 Pro is in partner testing while the team starts “its most ambitious pre-training run yet” for Gemini 4. The third model, Gemini 3.5 Flash Cyber, is a fine-tuned vulnerability finder restricted to governments and trusted partners via CodeMender pilots — a frontier lab deciding some coding capability doesn’t ship to everyone. Read more →

• Claude Code raised nested subagent depth from 1 to 3 — one week after the whole ecosystem agreed to cap it: v2.1.219 (July 24) makes Opus 5 the default Opus model with a 1M context window, adds a sandbox.network.strictAllowlist setting that denies non-allowlisted hosts without prompting, adds a DirectoryAdded hook, and quietly triples the depth at which subagents may spawn subagents. Last week three tools independently fenced that in. This week the fence moved — because the model behind it got good enough to trust deeper. Read more →

• Two prompt-engineering rules just inverted, according to the people who write Claude Code’s prompts: at the AI Engineer World’s Fair, Cat Wu and Thariq Shihipar said Claude Code’s system prompt has shrunk by 80% as models improved, that adding examples to a system prompt “is no longer best practice” for frontier models, and that negative constraints — “don’t do X” — now actively reduce output quality. Also from the same session: Claude Tag now lands 65% of the product engineering PRs on Anthropic’s product team, and agents never touch raw API keys thanks to credential injection. Read more →

• The most consequential Opus 5 number wasn’t on the benchmark chart: Boris Cherny points to page 73 of Anthropic’s system card — across evaluations and red-team testing, Opus 5 is the company’s least prompt-injectable model yet. Cherny’s own observation is that this got a fraction of the attention the coding scores did, which is backwards for anyone pointing an agent at untrusted content. Read more →

• GitHub Code Quality hit GA with no Copilot subscription required: now live on Enterprise Cloud and Team, it pairs CodeQL’s deterministic analysis with AI-assisted detection and Copilot Autofix to catch maintainability and reliability problems in the PR, with org-wide dashboards and coverage-based quality gates. GitHub’s number: teams resolve 67.3% of findings before merge. Read more →

📦 Releases & News

Codex ships 0.145.0 with Cursor and Claude Code settings import — then goes hands-free with full-duplex voice

OpenAI — July 21 and July 23, 2026

Two moves in three days, and the quieter one is the more strategic. Codex CLI 0.145.0 (July 21) added paginated thread history with efficient resume, Amazon Bedrock support with GPT-5.6 Sol as the default, a stabilized multi-agent V2 with configurable per-sub-agent models, and audio input for realtime conversations — plus the ability to import your settings directly from Cursor and Claude Code, which is less a feature than a switching-cost demolition. Then on July 23, release 26.715 put ChatGPT Voice into the desktop app, powered by GPT-Live — you can now “talk through work and coordinate tasks in Chat, Work, and Codex” by voice rather than typing. The same release let local projects span multiple related folders with a designated primary, which quietly fixes one of the more annoying constraints on pointing an agent at a real multi-repo workspace.

Copilot’s week: Opus 5 and Gemini 3.6 Flash land, the Linear agent goes GA, and Issues gets agent automation controls

GitHub Changelog — July 21–24, 2026

GitHub absorbed both of the week’s model releases almost on contact — Gemini 3.6 Flash across all Copilot tiers on July 21 with configurable reasoning effort and parallel tool use, and Claude Opus 5 on July 24, the same day Anthropic announced it. The more durable news is underneath: the Copilot cloud agent for Linear reached general availability on July 23, closing the loop from ticket assignment to open PR without leaving the tracker, the GitHub MCP Server adopted the next MCP specification, and agent automation controls for GitHub Issues entered public preview — the governance layer that has to exist before anyone lets an agent work a backlog unattended.

GitHub’s new Copilot dashboard stops counting seats and starts counting behaviour

GitHub Changelog — July 22, 2026

The metric that mattered used to be “how many licenses are active.” This dashboard replaces it with something far more uncomfortable and far more useful: it sorts your developers into three adoption phases — Code-first, Agent-first, and Multi-agent — plus a passive segment of people holding licenses they don’t use. Each cohort gets a card showing PR counts, merge velocity, lines of code, and headcount, and the whole thing rolls up into an “adoption multiplier” comparing throughput between your passive users and your engaged ones. Six-month trends included. If you have ever had to justify a Copilot renewal, this is the artifact you were missing.

OpenCode v1.18.4 and v1.18.5: adaptive thinking across four model providers, and a rewritten prompt input

OpenCode — July 20 and July 24, 2026

Two releases, both about keeping pace with how differently each provider now exposes reasoning. v1.18.4 (July 20) added adaptive-thinking controls for Kimi models, better reasoning-options handling, embedded terminal theme syncing, restored Azure Cognitive Services endpoint support, and a rewritten v2 prompt input for more reliable command, context, shell, attachment, and history interactions. v1.18.5 (July 24) extended the same work to Claude adaptive thinking, OpenAI response phases, and Mistral reasoning history, and brought the desktop client terminal transport, session actions, and event streaming.

Gemini CLI v0.52.0 stops letting the model “fix” your JSON

Gemini CLI — July 22, 2026

A small change with an outsized quality-of-life payoff: write_file and replace now bypass LLM correction entirely for JSON and IPYNB files, ending the class of bug where the model helpfully reformats a structured file into something subtly broken. The release also simplifies plan-mode policies to support relative paths, adds foundational triage-worker and GitHub Action handler modules for egress services, and finally surfaces a clear error when an account lacks Code Assist tier access instead of failing opaquely.

📚 Tutorials and Resources

Claude Code in Bun in Rust

Simon Willison — July 19, 2026

Simon Willison went looking for whether Claude Code actually runs the Rust-based Bun, and answered it the direct way: by taking the binary apart. Inside he found Bun v1.4.0 embedded, along with 563 Rust source filenames left in the build. His conclusion — “It looks like Bun in Rust is indeed being run in production across millions of different devices” — makes Claude Code the largest-scale production proof point the Bun-on-Rust rewrite has, which is a strange and fitting place for it to end up.

💡 Others

Here’s my uncomfortable bet on OpenAI

Ben Lorica, Gradient Flow — July 21, 2026

Lorica’s bet is that open-weight models eventually absorb most global AI spending on pricing pressure, efficiency, and control — and the reason the piece is worth your time is that he refuses to make it comfortable. He concedes that coding is precisely where US frontier models “ride a flywheel” with advantages “not closing quickly,” because serious users generate failure data and failure data trains better models. His prescription is architectural rather than ideological: build routers, OpenAI-compatible endpoints, and eval harnesses, and treat your prompts, code, and agent traces as strategic assets. In a week when both Anthropic and Google competed on cost per task rather than raw capability, that advice lands differently than it would have in January.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know that both Anthropic and Google spent this week competing on cost per task rather than capability — Opus 5 holding 5/5/25 while doubling Frontier-Bench, Gemini 3.6 Flash selling 17% fewer output tokens — and you know Claude Code raised nested subagent depth from 1 to 3 exactly one week after three tools agreed to cap it. You didn’t spend your weekend reading to know this. I did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,

Angel Llosa

LinkedIn · X · Medium

Questions, disagreements, or a story I missed? Just hit reply.

This week, compiled:

• The Money: Indian AI coding startup Emergent becomes a unicorn — $130M Series C at $1.5B, revenue up 70% in four months, selling “an engineering team in a box” to non-technical founders

• The Fork: Claude Code’s /fork now spawns full autonomous background sessions instead of in-session subagents — one of seven releases the tool shipped in six days

• The Privacy Scare: xAI open-sources Grok Build’s ~845K-line Rust codebase after users discovered it was uploading local directories — SSH keys included — to Google Cloud Storage by default

If you only read one thing this week: Moonshot AI shipped Kimi K3 with no keynote and no model card — just a quiet overnight flip of kimi.com to a 2.8-trillion-parameter model. Self-reported benchmarks put it ahead of Claude Opus 4.8 Max and GPT-5.5 High, at $3/$15 per Mtok — roughly half Opus 4.8’s per-task cost. Open weights land July 27, but the API is live now, and Moonshot is already calling it the first “open 3T-class model.” This is what the open-weight pressure on frontier pricing looks like when it stops being incremental. Read more →

Growing at 20% new subscribers per week.

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

Key Takeaways

• Bun’s AI-driven Rust rewrite cost ~$165,000 in API spend — and Zig’s creator called it “unreviewed slop”: Jarred Sumner ran up to ~64 parallel Claude Code instances over 11 days to port 535,496 lines of Zig into over 1,009,000 lines of Rust across 6,502 commits, versus an estimated 3 engineers working a full year unassisted. Andrew Kelley’s public rebuttal: “the tests pass” is not the same argument as reviewed. Read more →

• Emergent’s $130M Series C values it at $1.5B just over a year after launch, with 200,000+ paying customers and annualized revenue up 70% in four months — proof that the market for non-technical “vibe coding” platforms is scaling as fast as the developer-facing tools. Read more →

• xAI’s Grok Build privacy incident forced an unusual kind of transparency: after reports that the CLI uploaded entire local directories — including SSH keys and password databases — to Google Cloud Storage with retention on by default, xAI disabled the feature, flipped retention to off-by-default, and released the full ~845K-line Rust codebase under Apache 2.0. Read more →

• OpenCode’s v1.18.2 (July 15) stopped subagents from spawning nested subagents by default, adding a configurable subagent_depth limit — the same week Claude Code capped subagent spawns per session and Codex tightened dangerous-command detection. Three unrelated tools independently reached for the same fix within days of each other.

• Tetrate’s Agent Router Enterprise now brokers tokens automatically: set a budget cap once, and the gateway reroutes requests to an approved fallback model instead of failing the call when spend hits the ceiling. Tetrate’s own framing: “Agent token spend is the one line item the engineering organization can’t easily cap.” Read more →

• A benchmark that hit #1 on Hacker News found Claude Code loads ~33K tokens of system prompt and tool schemas before it even reads your prompt — versus ~7K for OpenCode, and up to ~75K in a realistic MCP/plugin-heavy setup. The catch: Claude Code’s batching of tool calls can still make it cheaper overall on multi-step tasks, so the fixed overhead isn’t the whole cost picture. Read more →

Subscribe now

📦 Releases & News

Codex CLI ships four releases (v0.144.2 → v0.144.6): security hardening and model-config fixes

OpenAI — July 13–18, 2026

Codex CLI’s week tracked the same theme running through Claude Code and OpenCode. v0.144.3 (July 13) added inline visualizations in Codex tasks, but the more consequential change came in v0.144.5 (July 16): improved dangerous-command detection, catching more forced rm forms and giving clearer rejection reasons when a command gets denied. v0.144.6 (July 18) closed out the week with refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrected context windows (272,000 tokens). Four tools, one week, one instinct: tighten what the agent is allowed to do on its own.

Cursor in Slack adds plan-sharing, multi-repo support, and cross-channel messaging

Cursor — July 17, 2026

Cursor’s Slack integration now shares a plan before it starts working, so you can redirect it early instead of discovering the wrong approach after the fact, and updates its status as it moves through each step. It can operate across multiple repositories in a single workspace with a one-click switcher, and — perhaps the most useful addition — it can now read from and post to other Slack channels and threads, pulling context from across a workspace rather than staying boxed into one conversation.

Security reviews land in the GitHub Copilot app

GitHub Changelog — July 14, 2026

The new /security-review slash command, now in public preview, scans in-flight code changes for injection, XSS, insecure data handling, path traversal, and weak crypto — with severity-scored, actionable remediation suggestions delivered before the code is ever committed. Another entry in the week’s broader story: as agents write more code unsupervised, the tooling around catching what they get wrong is catching up.

GitHub Copilot in Visual Studio’s June update: usage tracking, MCP trust baselines, and in-IDE PR review

GitHub Changelog — July 14, 2026

Six changes bundled into one release: a refreshed usage-tracking window with proactive billing alerts, MCP server configurations now validated against a trusted baseline at startup (any change requires explicit approval), general availability for the C++/MSVC modernization agent, edit suggestions that span the whole active file instead of just the area near your cursor, PR references inline in Copilot Chat, and a full in-IDE PR review flow — browse, comment, approve, complete — without leaving Visual Studio.

GitHub Copilot CLI ships six releases (v1.0.71-1 → v1.0.72-1): plan-mode hardening and MCP config persistence

GitHub — July 14–17, 2026

Another tool, the same instinct as the rest of this week’s shipping cycle: plan mode now hard-blocks built-in tool calls that would edit files or run mutating shell commands, closing a gap where “just planning” could still touch the filesystem. Smaller but useful fixes rounded out the week — GitHub MCP toolset/tool configuration now persists via settings.json, disabled skills are flagged in copilot skill list, invalid settings.json values now surface a specific warning instead of failing silently, and a bug that hung autopilot mode during long-running background processes was fixed (v1.0.71, July 16).

Zed 1.11.3: dedicated Staged/Unstaged Changes views and Git Graph improvements

Zed — July 15, 2026

Zed’s latest stable release adds dedicated Staged Changes and Unstaged Changes multibuffers, letting you review, stage, unstage, or restore individual hunks without leaving the editor — opened directly via git: view staged changes / git: view unstaged changes. Git Graph gets configurable columns and commit messages rendered as Markdown, and the project symbols picker adds a preview pane, alongside 60+ other fixes.

This was the week the ecosystem quietly agreed that more autonomy needs more fencing. Claude Code capped how many subagents and web searches a session can spawn on its own and patched a run of permission-check bypasses in the same breath it shipped /fork for full background sessions. OpenCode, days later, stopped subagents from spawning their own subagents by default. Codex tightened what counts as a dangerous command, and GitHub Copilot CLI closed the gap where “plan mode” could still quietly edit files. Four tools, no coordination, the same fix — because agents that can now run unattended for longer need harder limits on what “unattended” is allowed to mean. And in case anyone needed the counterargument spelled out, Zig’s creator spent the week arguing publicly that Bun’s own million-line, $165,000, AI-generated Rust rewrite is exactly the kind of code those guardrails exist to catch — reviewed or not. Meanwhile Kimi K3 landed without ceremony at half of Opus’s price, Emergent proved the non-technical end of this market is scaling just as fast, and a privacy scare at xAI turned into the most detailed public look yet at what a frontier coding agent’s codebase actually contains.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know why three unrelated coding agents all shipped subagent-spawn limits in the same six days, and you know Kimi K3 undercut Opus 4.8 pricing by half without so much as a press release. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,

The Vibe Coding Team.

• The Data: The share of developers who stopped manually reviewing AI-generated commits jumped from 10% to 40% in a single month

• The Money: Bun rewrote a core subsystem in Rust with AI in 11 days for $165K in tokens — work estimated at a year for a small human team

• The Trend: Token budgets, not headcount, are the flexible line item — one security firm cut LLM spend 59–70% by fixing its cache hit rate

If you only read one thing this week: Three frontier labs shipped coding-capable models within a single 48-hour window. OpenAI took GPT-5.6 to general availability on July 9 as a three-tier family — Sol (flagship), Terra (balanced), and Luna (fast/cheap) — landing in ChatGPT, Codex, and the API, one day after xAI’s Grok 4.5 and the same day as Meta’s Muse Spark 1.1. If your team standardized on a model last quarter, this is the week that decision reopened. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones
actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you
arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework
for the conversation that always comes after “we should use AI more”: how to
actually move an organization that didn’t ask to be moved. Included with every
subscription.

Subscribe now

Key Takeaways

• AI code review discipline is collapsing faster than anyone predicted: median Cursor users generate ~700 lines of code weekly while the top 1% generate 30,000–40,000 lines (roughly 45 median developers’ worth) — and the share of developers who no longer manually review AI-generated commits quadrupled to 40% in March 2026 alone. Read more →

• “AI did the rewrite” now has a price tag: Bun’s AI-assisted Rust rewrite of a core subsystem took 11 days and $165K in tokens, versus an estimated year for a small human team — one of the first public case studies with hard cash numbers instead of anecdote. Read more →

• Caching is the cheapest engineer you’ll hire this year: security firm ProjectDiscovery raised its cache hit rate from 7% to 84% and cut LLM spend 59–70% — while Uber handed AI tools to 5,000 engineers and burned its entire 2026 AI budget by April. Read more →

• xAI trained Grok 4.5 alongside Cursor’s data flywheel — a rare public confirmation of a frontier lab and an agentic IDE co-training a model, blurring the line between model provider and tool vendor. Read more →

• Meta finally put a price on its models: Muse Spark 1.1 is Meta’s first paid model — 1M-token context at $1.25/$4.25 per million tokens, roughly a quarter of Anthropic and OpenAI’s rates, launched via a new public-preview Meta Model API. Read more →

📦 Releases & News

Claude Code Ships Six Releases in a Week (v2.1.202 → v2.1.207)

Claude Code enabled auto mode by default on Bedrock, Vertex AI, and Foundry (opt-out via disableAutoMode), bumped Bedrock to Opus 4.8, and reworked agent-view rows to show classifier-written headlines instead of raw tool-call text. The fixes matter too: a terminal freeze on long streamed outputs, a consent-dialog bypass on remote managed settings, and an auto-updater bug that clobbered custom launcher symlinks. If you run Claude Code through a cloud provider, the new default is worth a deliberate look before your next session.

OpenAI Codex Enables Remote Plugins by Default, Joins ChatGPT Desktop

Across CLI 0.143.0–0.144.1 (July 6–9), Codex turned on remote plugins by default with a richer catalog and proxy-aware auth on macOS/Windows, added Bedrock GPT-5.6 models, and shipped configurable rollout token budgets that abort a turn when the budget is exhausted — a native cost-control primitive landing directly in the agent. Codex also moved into the ChatGPT desktop app, letting users edit code and review PR diffs without leaving ChatGPT.

OpenCode Adds Code-Mode MCP Adapter, Fixes Copilot Pricing Crash

Across four releases (July 7–9, through v1.17.18), OpenCode shipped a code-mode MCP adapter, a v2 free-model selector for faster access to more providers, and broad desktop/TUI workflow fixes. The most instructive item is a crash caused by GitHub Copilot returning models with a zero billing batch size — a reminder that provider-side pricing metadata bugs can break your agents downstream, whether or not you touched anything.

Gemini CLI v0.50.0 Adds Tool Registry Discovery

The stable v0.50.0 release (July 8) introduced tool registry discovery for automatic detection and registration of tools, plus CI pipeline safeguards to stop bad npm releases from shipping. A same-day preview (v0.51.0-preview.0) hardened path-resolution security for at-reference files — a quiet but meaningful fix given how much context flows through file references in agent sessions.

GitHub Copilot in VS Code Ships Agentic Browser Tools, Parallel Sessions

VS Code’s Copilot agent can now browse and validate web apps directly (agentic browser tools are GA), run multiple parallel agent sessions with separate chats, and expose session-level and subagent-level credit cost tracking — cost observability arriving as a first-class IDE feature, not an afterthought. A new Language Models editor also lets you discover and install model-provider extensions on the fly.

Cursor 3.11 Adds Side Chats and Agent Transcript Search

Cursor 3.11 (July 10) lets you branch off a side chat (via /side, /btw, or a plus button) to explore tangents without derailing the main agent conversation, and adds full-text search across past agent transcripts. New cloud-agent hooks (beforeSubmitPrompt, afterAgentResponse, afterAgentThought) give teams programmatic visibility into agent runs — the audit layer enterprises keep asking for.

Zed Adds llama.cpp as a Model Provider, Day-One GPT-5.6 Support

Zed 1.10 (July 8–10, three releases) added llama.cpp as a first-class language model provider — a notable nod to local and self-hosted model workflows — and unified LLM providers, external agents, and MCP servers into a single settings editor. Within 24 hours of GPT-5.6’s release, Zed shipped support for Sol and Terra, with Luna pending third-party API access.

💡 Others

Tech Jobs Market in 2026, Part 3: Hiring Managers & Job Seekers

The Pragmatic Engineer examines how AI adoption is reshaping engineering hiring: which AI-related roles are hottest, and why conditions remain tough for engineering leaders despite strong demand for AI skills. Worth reading if you’re on either side of the hiring table — the skills premium is real, but it isn’t landing where most job seekers expect.

The arc of this week is simple: the models multiplied, and the money got named. Three frontier launches in 48 hours would have been the whole story a year ago — this week it shared the stage with a $165K rewrite invoice, a cache hit rate that saved more budget than a layoff round, and 40% of developers waving AI commits through unreviewed. The tools are no longer the bottleneck. The economics and the discipline are.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes
everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped
three viral threads that turned out to be nothing. You know which of the three new frontier models actually changes your team’s pricing math, and you know that the scariest number this week wasn’t a benchmark — it was 40% of developers no longer reviewing AI commits. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and
everyone building or managing software in the age of AI.

Clean code and positive vibes,

The Vibecoding Team.

• The Big Model: Claude Sonnet 5 launches with performance approaching Opus 4.8 at introductory $2/$10 per Mtok — now the default model inside Claude Code v2.1.197

• The Agent: Claude Code background subagents run by default and auto-commit, push, and open draft PRs when they finish — your terminal now manages its own ticket queue

• The Ecosystem: Kimi K2.7 Code becomes the first open-weight model in GitHub Copilot’s model picker, cracking open a platform that had been closed-weight since day one

If you only read one thing this week: The US government lifted export controls on Fable 5 and Mythos 5 on June 30. What matters more than the return itself is what Anthropic published alongside it: a cross-industry jailbreak severity framework (CJS-0 to CJS-4) co-developed with Amazon, Microsoft, and Google — the first attempt to give every lab, researcher, and regulator a shared language for how dangerous a jailbreak actually is. This is what AI safety governance looks like when it starts to get real. Read more →

Subscribe now

Key Takeaways

• Claude Sonnet 5 is the best value Anthropic has ever shipped: near-Opus 4.8 performance, introductory pricing of $2 input / $10 output per Mtok through August 31, available across all plans and inside Claude Code as the new default. Read more →

• Claude Code shipped 6 releases in 7 days, with v2.1.198 being the pivotal one: background subagents now run in parallel by default, auto-committing, pushing, and opening draft PRs when they finish. Claude in Chrome also hit general availability. Six releases, one week, fundamentally different terminal. Read more →

• Kimi K2.7 Code is the first open-weight model in GitHub Copilot’s model picker — available now for Pro, Pro+, and Max; requires admin opt-in for Business and Enterprise. It’s one switch, but it signals a structural shift in how GitHub thinks about model diversity on the platform. Read more →

• Copilot browser tools reached GA in VS Code: agents can now navigate, click, type, take screenshots, and feed live web content directly back into chat — with user-controlled tab sharing and fully isolated sessions that have no access to your cookies or browsing history. Read more →

• Claude Science launched as an AI workbench for researchers with 60+ pre-configured skills across genomics, proteomics, and cheminformatics, a reviewer agent that checks citations and flags errors, and scaling from laptops to HPC clusters. Applications open through July 15 for up to $30K in credits per project. Read more →

• Claude Opus 4.8 and Haiku 4.5 are now generally available in Microsoft Foundry on Azure, with Azure-native billing, authentication, networking, and US data residency. Eligible customers with a Microsoft Enterprise Agreement can draw Claude usage down against their existing Azure commitment. Read more →

• Microsoft Research’s Memora cuts AI agent context token usage by up to 98% using a harmonic memory architecture that decouples what is stored from how it’s retrieved — achieving state-of-the-art memory accuracy (87.4% on LongMemEval) while storing half the memory entries of Mem0 per conversation. As agentic coding tools make 10–100x more LLM calls than chatbots, techniques like Memora are becoming the FinOps layer for AI teams. Code is on GitHub; paper presented at ICML 2026. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription FOR FREE.

Subscribe now

📦 Releases & News

More details on Fable 5’s cyber safeguards and our jailbreak framework

Anthropic — July 2, 2026

Two days after Fable 5’s return, Anthropic published the technical details most developers will actually want. The model’s cybersecurity classifier now divides use cases into four tiers — from outright prohibited (ransomware, malware, exfiltration) to benign IT work (secure coding, patch management) — with the classifier accepting higher false-positive rates in favor of safety margin. Separately, the Cyber Jailbreak Severity (CJS) framework scores any reported jailbreak across four axes: capability gain, breadth of enabling offensive tasks, ease of weaponization, and discoverability, producing a score from CJS-0 (Informational) to CJS-4 (Critical). The framework is co-developed with Amazon, Microsoft, and Google, and is designed to be adopted across labs — not just Anthropic.

Claude Code v2.1.200 — Manual permissions as the new default

Anthropic — July 3, 2026

The quietest but most consequential change in the week’s shipping sprint: Claude Code’s default permission mode is now “Manual” across CLI, VS Code, and JetBrains. AskUserQuestion dialogs no longer auto-continue (you can re-enable idle timeout via /config). The release also fixed background session failures after sleep/wake cycles, resolved subagents cut off by rate limits returning empty results instead of errors, and improved terminal rendering under tmux 3.4+. For teams running unattended agents, these reliability fixes matter as much as any feature.

OpenCode v1.17.12 and v1.17.13 — Sonnet 5 adaptive thinking and new model picker

OpenCode — June 30 and July 1, 2026

OpenCode shipped two releases in two days to respond to the Sonnet 5 launch. v1.17.12 enables adaptive thinking for Claude Sonnet 5 alongside improved MCP OAuth (refresh-token scope, better error surfaces) and token/cost totals visible in session context. v1.17.13 adds a searchable v2 model picker with model management flow, session tab hover previews showing project, path, branch, and server at a glance, and streamlined WSL server setup. Taken together: OpenCode turned around Sonnet 5 compatibility in under 24 hours of the model going live.

Google launches Nano Banana 2 Lite and Gemini Omni Flash

Google Blog — June 30, 2026

Two developer-facing models shipped the same day as Sonnet 5. Nano Banana 2 Lite is Google’s fastest and cheapest image model — 4-second generation times at $0.034 per 1K images, available immediately in Google AI Studio and the Gemini API. Gemini Omni Flash enters public preview as a natively multimodal video generation model supporting conversational editing from text, image, and video inputs at $0.10 per second of output. Both are accessible through the Gemini API, Google AI Studio, and the Enterprise Agent Platform starting today.

Copilot CLI no longer needs a PAT in GitHub Actions

GitHub Changelog — July 2, 2026

A simple change that removes significant operational overhead: GitHub Copilot CLI in Actions now authenticates with the built-in GITHUB_TOKEN, requiring only the copilot-requests: write permission. Personal access tokens — long-lived credentials that create security exposure and rotation headaches at scale — are no longer needed. AI credits bill directly to the organization. If your CI uses Copilot CLI today, this is a drop-in improvement with no behavior change.

Introducing Laguna XS 2.1

Poolside — July 2, 2026

As frontier API costs become harder to justify at scale, the open-weight option keeps getting better. Poolside’s Laguna XS 2.1 is a 33B-parameter MoE model that runs on a single GPU, achieves 63.1% on SWE-bench Multilingual (+5.4 points over its predecessor), and is available free on Hugging Face and OpenRouter under the fully permissive OpenMDW-1.1 license. For teams that need local agentic coding — private codebases, air-gapped environments, or simply more predictable infrastructure costs — Laguna XS 2.1 is a credible alternative to frontier API tools, deployable via Ollama, llama.cpp, vLLM, or TensorRT-LLM, with a paid hosted tier at $0.10/$0.20 per Mtok if you prefer not to self-host.

💡 Others

Copilot agent session streaming is now in public preview

GitHub Changelog — July 2, 2026

Enterprise visibility into agent activity just became possible. GitHub Enterprise Cloud customers can now stream Copilot agent session data — every prompt, response, and tool call — across all Copilot clients to a SIEM, event collector, or Microsoft Purview endpoint. Alternatively, pull up to 48 hours of records on-demand via REST API. As agents become standard in enterprise workflows, having an audit trail for what they did and why is increasingly non-negotiable — this is the first version of that answer from GitHub.

Upcoming deprecation of Gemini 2.5 Pro and Gemini 3 Flash in GitHub Copilot

GitHub Changelog — July 2, 2026

Action required if your team uses Gemini models via Copilot: GitHub has announced the upcoming deprecation of both Gemini 2.5 Pro and Gemini 3 Flash from the Copilot model picker. Teams relying on either should plan a migration to Gemini 3 Pro or an alternative in the picker before the deprecation date. Given the simultaneous announcement that Kimi K2.7 Code is now available, this week’s GitHub Copilot model story is really about the picker becoming a more dynamic surface — older models out, new open-weight options in.

This was a week where the word “default” did a lot of work. Sonnet 5 is now the default model in Claude Code. Background agents are now the default operating mode. Manual permissions are now the default safety posture. Anthropic shipped six releases in seven days, and each one quietly changed what it means to open a terminal. Meanwhile, Fable 5 came back — and what came with it was the first real attempt at a shared industry standard for measuring how dangerous a jailbreak is. A government suspended a model for three weeks and a cross-company safety framework came out the other side. And in the background, a Microsoft Research team published a paper showing you can cut your agent’s context token bill by 98% — because as your agents get more capable, the question of what they cost per task is no longer optional.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know Fable 5 is back globally, why it was suspended, and what the CJS framework means for every lab that ships frontier models. You know Claude Sonnet 5 benchmarks near Opus 4.8 at less than half the price, that Claude Code is now opening your pull requests in the background, and that Microsoft Research just published a technique to cut your agent’s token bill by 98%. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel.

• The Big Story: Anthropic launches Claude Tag — a persistent AI teammate inside Slack that 65% of Anthropic’s own product team uses to write code, with multiplayer collaboration, async task delegation, and admin-controlled tool access.

• The Release: Claude Code shipped five versions in one week, cutting streaming CPU usage by 37%, adding /rewind to undo a /clear, and routing all shell commands through auto-mode classification.

• The Efficiency Win: GitHub’s Copilot agentic harness, benchmarked against Claude, GPT-5.4, and GPT-5.5 on five industry benchmarks, matches model-vendor harnesses on task completion while consuming fewer tokens — the platform layer is quietly becoming the competitive moat.

If you only read one thing this week: OpenAI previewed GPT-5.6 — three models named Sol, Terra, and Luna — but almost nobody can use them yet. For the first time, a US executive order required OpenAI to coordinate with the federal government before releasing a frontier AI model to the public. Only ~20 trusted partner organizations have access. Sol handles complex coding and security research; Terra does the same work as GPT-5.5 at half the cost; Luna is the fastest and cheapest for everyday tasks. The restriction isn’t just a regulatory footnote — it’s the first time the US government has formally inserted itself into the frontier model release cycle. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

Key Takeaways

• Microsoft’s first own-brand coding model is now GA on Copilot. MAI-Code-1-Flash, built entirely in-house by Microsoft AI, became generally available for Copilot Business and Enterprise on June 26. It’s optimized for high-volume iterative agentic workflows — fast and low-latency by design, billed at usage-based rates. Admins need to manually enable the policy. This marks Microsoft’s first step from model reseller to model maker inside its own developer platform. Read more →

• Codex Remote is generally available: AI coding you can approve from your phone. As of June 25, developers can start or continue Codex sessions on a connected Mac or Windows host and monitor and approve actions from an iOS or Android device. The release adds authenticated QR pairing between mobile device and host, and ships a new DigitalOcean plugin that provisions a cloud Droplet as a remote workspace from inside the Codex app. Read more →

• GitHub Actions can now run steps in parallel. Four new keywords landed on June 25: background: true starts a step asynchronously; wait/wait-all blocks until specific background steps finish; cancel gracefully stops a background step; and parallel wraps multiple steps in concurrent execution with automatic waiting. Common use cases include running parallel builds, spinning up services while other steps proceed, and non-blocking telemetry uploads. Read more →

• Free and Student Copilot users can no longer choose their AI model. Starting June 24, GitHub replaced manual model selection on Free and Student Copilot plans with automatic routing only. “Auto” dynamically picks the best available model per task from across the OpenAI, Anthropic, Google, and Microsoft model families — but which models are in the pool can change without notice. GitHub also retired the “(Preview)” label from all Microsoft-released models as part of the same update. Read more →

• 5x more pull requests, but code review is collapsing. The Pragmatic Engineer’s Gergely Orosz published a June 23 analysis of what AI productivity data actually means at scale. Teams using agents ship 5x more PRs and generate 2.5x more code per developer than two years ago — but Cursor’s own data shows AI-generated code is increasingly being merged without any human review, a trend that accelerated sharply in February 2026. Meta reallocated engineers from security and quality to AI training tasks, then suffered a high-profile breach. Orosz’s conclusion: as output volume becomes cheap and automatic, “taste” — deliberate human judgment about what to ship — becomes the scarcest skill in the room. Read more →

📦 Releases & News

GitHub Desktop 3.6: Worktrees and Deeper Copilot Integration
GitHub Desktop 3.6, released June 26, adds native Git worktree support so developers can work across multiple branches simultaneously without stashing or re-cloning — a change that matters most when running parallel AI agent sessions on the same repository. Copilot now generates commit messages that respect custom instructions from .github/copilot-instructions.md and AGENTS.md, and provides AI-assisted merge conflict resolution. A model picker lets users choose from all Copilot-accessible models or bring their own key.

GitHub Copilot for Jira Is Now Generally Available
Copilot for Jira exited public preview on June 25 with three GA additions: real-time progress monitoring of the coding agent directly inside a Jira issue, the ability to give follow-up instructions in the Jira chat panel after the agent creates a draft PR (so it keeps refining the same PR instead of creating a new one), and streamlined setup requiring fewer configuration steps. Preview-phase features — model selection, Confluence context via MCP, custom agents, space-level guidance, and review request notifications — are all now generally available too.

Claude as Agent Provider in JetBrains IDEs — Public Preview
As of June 22, JetBrains IDE users can select Claude as their Copilot agent provider by pointing the IDE to the Claude Code CLI path. Important caveat for the preview: Claude currently runs in bypass permissions mode, meaning all file edits and tool calls are auto-approved without confirmation. Configurable permissions are on the roadmap. The release also ships organization and enterprise agent support, a model picker, and a per-turn AI credits indicator.

Cursor 3.9: Customize Page, Plugin Marketplace, and Team Workspaces
Cursor 3.9 landed on June 22 with a new Customize page that centralizes management of plugins, skills, MCPs, subagents, rules, commands, and hooks across user, team, and workspace scopes. A marketplace leaderboard surfaces the most-used extensions across your team for one-click installation. Plugin canvases ship prebuilt templates: Hex Canvas for data visualizations, Atlassian Canvas for real-time project visibility. Team Marketplaces now import plugin repositories from GitLab, Bitbucket, and Azure DevOps — not just GitHub.

OpenCode v1.17.11: Session Snapshots and Revert Controls
OpenCode’s June 25 release adds the ability to roll a session back to any earlier message, including all associated file changes — a safety net that makes it easier to experiment aggressively and recover cleanly. Desktop improvements add Chrome-style tab cycling and draggable tabs. Bug fixes cover prompt draft attachment persistence and titlebar tab consistency. This brings session-level undo to what has been a rapid release cadence from the sst/opencode team.

OpenCode v1.17.10: MCP Server Instructions and Managed Provider Integration
The June 24 OpenCode release injects MCP server instructions directly into the session context, giving the model visibility into what each connected MCP server can do without requiring a separate lookup. It also adds support for Opencode-managed provider integrations, a configurable keybind for the diff viewer, mobile bottom navigation in desktop mode, and collapsible server sections in the sidebar.

The week’s theme is consolidation under pressure. GPT-5.6 lands behind a government gate. Claude Tag turns AI into a team member with a Slack handle. GitHub ships its own model. Harnesses are being benchmarked like infrastructure. And Orosz reminds us that the PRs-per-developer metric, taken alone, is a trap. The tools are genuinely getting better. The question of who’s watching the output isn’t going away.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know GPT-5.6 exists but why almost nobody can use it yet, and you understand why a week with five Claude Code releases and a new Microsoft model actually matters for how your team works. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,

The Vibe Coding Team.

• The Big Story: Google shut down Gemini CLI on June 18 — 103K GitHub stars, 6,000 community contributors, and every CI/CD pipeline that depended on the gemini command stopped receiving responses overnight. The replacement, Antigravity CLI, is a closed-source Go binary with no 1:1 feature parity at launch.

• The Automation: Codex shipped Record & Replay on macOS — show it a workflow once, and the language model converts the demonstration into a reusable, inspectable, editable skill. No scripting, no rule-based heuristics. Cursor matched the moment with /automate, Slack emoji triggers, and computer use for cloud agents — all on the same day.

• The Safety Net: Claude Code v2.1.183 now blocks destructive git commands in auto mode — git reset --hard, git checkout -- ., git clean -fd, and git stash drop are all rejected unless you explicitly ask for them. Same goes for terraform destroy, pulumi destroy, and cdk destroy.

If you only read one thing this week: A security team called Tenet published a new attack class called “Agentjacking” — and the results are uncomfortable. One fake Sentry error report, injected through a publicly accessible DSN, was enough to hijack AI coding agents into running attacker-controlled code on developer machines. Claude Code, Cursor, and Codex all fell for it with an 85% success rate. The attack bypasses every traditional security control because every action in the chain is technically authorized. Tenet found 2,388 organizations with exposed Sentry DSNs during their validation window. Sentry acknowledged the issue the same day it was disclosed — and declined to fix it, calling the attack class “technically not defensible” at the platform level. Read more →

Subscribe now

Key Takeaways

• Gemini CLI’s shutdown is the open-source trust story of the year: Google accepted 6,000+ merged pull requests under Apache 2.0 with a CLA that granted perpetual, irrevocable rights — then used that code to build Antigravity CLI, a closed-source successor. Enterprise users with Gemini Code Assist Standard/Enterprise licenses keep access, but every free and Pro user lost their CLI on June 18 with no open-source alternative. Read more →

• Claude Code Artifacts turn coding sessions into live, shareable web pages: Artifacts publish to a private URL on claude.ai, update in real-time as the session works, support version history, and enforce org-only visibility via strict CSP — 16 MiB per page, no external network requests. PR walkthroughs, dashboards, and incident pages that stay current without anyone manually updating them. Read more →

• Codex Record & Replay teaches the agent by demonstration instead of description: Show Codex a workflow — uploading a YouTube video, filing an expense report — and it generalizes the steps into a reusable skill using the language model, not rule-based heuristics. The skill is inspectable and editable. Available on macOS for Plus, Pro, Business, Enterprise, and Edu tiers. Read more →

• Claude Design now imports your design system from GitHub and round-trips code with Claude Code: The /design and /design-sync commands create a bidirectional pipeline between design and code — the first time a major AI tool has closed the design-to-implementation loop without screenshots or manual rebuilds. Nine new export destinations (Adobe, Canva, Lovable, Replit, Vercel) and shared token limits across Claude products complete the overhaul. Read more →

• Cursor 3.8 makes automations triggerable from Slack and GitHub — with computer use for cloud agents: The /automate skill lets you describe an automation in plain language, but the real addition is five new GitHub triggers (issue comments, PR reviews, review thread resolution, Actions completion) and Slack emoji triggers that let teams kick off agent workflows without opening the IDE. Read more →

Growing at 20% new subscribers per week.

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Devin Desktop v3.2.16: Plugin System for Devin Local

Devin Local is becoming an extensible platform. The June 16 release introduces a plugin system (preview, opt-in for enterprises) that lets teams add custom tools, skills, and integrations to their local agent. Subagents can now call MCP tools directly — a meaningful expansion of what Devin Local can reach without routing through the cloud. Teams also gain CLI permission scopes to enforce terminal allow/deny lists, giving enterprises the governance layer they’ve been requesting since the Windsurf-to-Devin transition.

Codex CLI v0.141.0: Encrypted Relay Channels and Plugin Marketplace

The June 18 stable release ships 87 changes with a clear security theme: remote executors now communicate through authenticated, end-to-end encrypted Noise relay channels, and cross-platform execution maintains native working directories across system boundaries. The plugin marketplace gets more informative automation — codex plugin marketplace list --json now includes each marketplace source, and plugin lists return from cached catalogs before refreshing in the background. Combined with tool search caching and reduced memory consumption, this is the most infrastructure-focused Codex release in months.

Claude Code v2.1.183: Auto Mode Safety Blocks Destructive Commands

The June 19 release is all about preventing the agent from doing things you didn’t ask for. Destructive git commands — git reset --hard, git checkout -- ., git clean -fd, git stash drop — are now blocked in auto mode unless you explicitly request them. The same goes for git commit --amend on commits the agent didn’t make, and terraform destroy, pulumi destroy, and cdk destroy. Also adds model deprecation warnings in print mode and a new attribution.sessionUrl setting to omit claude.ai session links from commits and PRs.

Tenet Security Launches with $6M Seed Round and Agentjacking Disclosure

The same day Tenet Security published its Agentjacking research, the company emerged from stealth with a $6M seed round led by The Westly Group (early SentinelOne investor). Founded by ex-Cisco AI Defense researchers, Tenet’s core technology — patent-pending “Agent-side Simulation” — predicts an AI agent’s likely next actions before they execute in production. The timing is deliberate: a new attack class and a funded startup to address it, launched on the same day.

OpenCode v1.17.7 / v1.17.8 / v1.17.9: MCP Compatibility Push

Three OpenCode releases in one week, all focused on making the MCP ecosystem work more reliably. v1.17.7 (June 14) fixes plugin client server reuse and adds ACP shell tool visibility. v1.17.8 (June 17) makes OpenAI-compatible providers accept MCP tool schemas that previously failed validation and fixes Cloudflare AI Gateway key handling. v1.17.9 (June 21) honors configured agent step limits and improves prompt caching. Small individually, but collectively they reflect OpenCode’s bet on being the most interoperable agent in the market.

📚 Tutorials and Resources

Kiro CLI V3 Early Access: Spec-Driven Development Comes to the Terminal

Kiro CLI V3 launches in early access on June 17, running alongside existing 2.x installations via kiro-cli --v3. The new version brings the unified agent harness used across IDE and Web to the terminal for the first time — with spec-driven development, a capability-based permissions model, enhanced hooks with a standalone file format, and tag-based agent configuration. Two days later (June 19), Kiro Web shipped Automations: schedule recurring work with GitHub/GitLab repos, up to 5 schedules per automation, each running autonomously in a sandbox and opening PRs when done.

Claude Code v2.1.181: Inline Config, Bun 1.4, and Streaming Overhaul

The June 17 release adds /config key=value syntax for changing settings without leaving the conversation — a small but meaningful friction reduction for power users who constantly toggle features. Under the hood, the bundled Bun runtime upgrades to 1.4, streaming now delivers long paragraphs line-by-line instead of waiting for the first break, and sandbox.allowAppleEvents opens macOS automation capabilities in sandboxed sessions. A new CLAUDE_CLIENT_PRESENCE_FILE environment variable lets server setups suppress mobile notifications.

💡 Others

Fable 5 Tops Terminal-Bench 2.1 — But No One Can Use It

The Terminal-Bench 2.1 leaderboard updated on June 17 with Claude Fable 5 entries — and the gap is striking. Fable 5 hit 88.0%, the first model to break 85% on this benchmark, sitting 4.6 points ahead of GPT-5.5 at 83.4%. Claude Code + Fable 5 scored 83.1%, Terminus 2 + Fable 5 at 80.4%. The irony: the model that topped the leaderboard remains suspended under a US export-control directive since June 12, meaning no one can actually use it.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know that Google killed Gemini CLI and 6,000 contributors’ trust overnight, that a single fake Sentry error can hijack every major coding agent on the market, and that your Claude Code sessions can now become live web pages your team watches update in real time. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel.

• The Big Story: GitHub Copilot’s switch to usage-based billing went live on June 1 — and developers are reporting 10x–50x cost increases overnight, with agentic coding sessions eating through a month of credits in a single sitting.

• The New Default: Windsurf retired after one year and shipped as Devin Desktop — Cascade replaced by a Rust-rewritten agent, Agent Command Center as the new home screen, and open ACP protocol letting any agent run alongside Devin in the same workspace.

• The Platform Bet: GitHub shipped Agentic Workflows in public preview, GitHub Copilot Sandboxes, and the Copilot SDK GA all in the same week — making GitHub the platform layer for running agents, not just the host for their output.

If you only read one thing this week: Claude Fable 5 launched on June 9 as the first publicly available Mythos-class model, and the benchmark gap is significant: 80.3% on SWE-bench Pro versus 69.2% for Opus 4.8 and 58.6% for GPT-5.5. The real number isn’t on a leaderboard, though — it’s in a Stripe engineering report. Fable 5 completed a codebase-wide migration of a 50-million-line Ruby codebase in one day. The same task had been estimated at two months of team work. At $10/$50 per million input/output tokens (2× Opus 4.8), it was free on all paid Claude plans through June 22 — but the story took a sharp turn on June 12, when a US government export-control directive forced Anthropic to suspend access to Fable 5 and Mythos 5 worldwide, for every customer inside and outside the United States, citing national-security concerns over a potential jailbreak of Fable’s safeguards. Anthropic disputes the decision — calling the vulnerability a “narrow, non-universal jailbreak” no different from competing models — but is complying. All other Anthropic models remain available. The most capable coding model of the year was publicly accessible for three days. Read more →

Key Takeaways

• GitHub Copilot’s billing shock is the first crack in the “unlimited AI for a flat fee” model: Developers report projections jumping from $29/month to $750 and from $50/month to $3,000 — agentic coding sessions routinely hit $30–$40 each, blowing through a Pro user’s $10 monthly credit ceiling in one sitting. Code completions remain unlimited, but the economics of agentic workflows at scale have fundamentally changed. Read more →

• GitHub Agentic Workflows in public preview lets you define CI automation in plain Markdown: The system compiles natural-language workflow descriptions into standard GitHub Actions YAML, supports issue triage, CI failure analysis, and code simplification, and runs agents in sandboxed containers with an Agent Workflow Firewall — no personal access token required, GITHUB_TOKEN is enough. Read more →

• The Copilot SDK GA turns GitHub’s agentic engine into an embeddable building block: Available across six languages (Node.js, Python, Go, .NET, Rust, Java), it exposes custom tool registration, MCP server integration, lifecycle hook interception, and OpenTelemetry tracing — meaning any developer tool can now embed Copilot’s agentic reasoning layer, not just use it through GitHub’s surfaces. Read more →

• OpenCode sub-agents can now spawn sub-agents up to 5 levels deep: This mirrors Claude Code’s own nested sub-agent architecture (also shipping this week in v2.1.172) — a convergence across tools toward multi-agent pipeline depth as the default architectural pattern for complex coding tasks. Read more →

• Cursor Bugbot is now 3x faster, finds 10% more bugs, and costs 22% less: The combination of speed, quality, and cost moving simultaneously in the same direction is unusual — Bugbot previously felt like a slow tax you paid for safety, and the /review command now makes it a natural pre-push habit. Read more →

• Kiro’s spec-driven workflow now works on GitLab and in the browser: Teams that live in GitLab can now plan work as reviewable requirements and task files in Kiro Web before the agent writes a line of code — extending a Spec-First approach that was previously limited to GitHub-connected repositories. Read more →

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Windsurf Is Now Devin Desktop

Cognition retired the Windsurf brand on June 2 and shipped Devin Desktop as a standard over-the-air update. Three things materially changed: Cascade is replaced by Devin Local — a Rust rewrite that is up to 30% more token efficient and supports subagents; the Agent Command Center is now the default screen when you open the app instead of the code editor; and the product ships with open Agent Client Protocol support, letting Claude Agent, Codex, and OpenCode run alongside Devin in the same workspace. Existing Windsurf plans, pricing, and extensions carry over without changes. Cascade remains available until July 1, 2026.

OpenAI Launches Codex for Every Role: Role-Specific Plugins and Sites

OpenAI launched six role-specific Codex plugins (data analytics, creative production, sales, product design, equity investing, investment banking) alongside Codex Sites — the ability to create and host interactive web apps and dashboards that you can share via URL. Codex now has 5 million weekly active users, up 6x since the desktop app launched in February. Knowledge workers, who make up 20% of users but are growing 3x faster than developers, are the intended primary audience for this expansion. This marks a deliberate move from Codex as a coding tool to Codex as a cross-functional work platform.

Claude Code v2.1.169: Safe Mode and /cd Command

Claude Code ships two significant quality-of-life changes on June 8–9: --safe-mode disables all customizations while preserving authentication, giving you a clean diagnostic environment when a plugin or configuration breaks things; and /cd moves your session to a new working directory without breaking the prompt cache. The same release adds post-session hooks, fixes Up/Down arrow history navigation for long inputs, and tightens enterprise MCP policy enforcement on reconnect.

Claude Code v2.1.172: Sub-Agents Spawn Sub-Agents

The June 10 release enables sub-agents to spawn their own sub-agents up to 5 levels deep, unlocking recursive multi-agent pipeline architectures inside a single Claude Code session. Also in this release: a plugin marketplace search bar in /plugin, model attribute added to OTEL metrics for per-model cost attribution, and performance improvements that reduce idle CPU usage. Background agent directory isolation is also fixed, preventing cross-agent state leakage.

GitHub Copilot App Opens to All Subscribers with Canvases

The GitHub Copilot app technical preview expands to all Pro, Pro+, Business, and Enterprise subscribers. The focal feature is Canvases — bidirectional work surfaces where agents write to actual work objects (plans, PRs, checklists) rather than chat messages, making progress visible and editable in real time. Voice input via on-device speech recognition is also included, and cloud-based agent sessions let you hand off from CLI into the app interface.

Cloud and Local Sandboxes for GitHub Copilot in Public Preview

GitHub ships two sandbox types for Copilot agents: local sandboxes (built on Microsoft MXC technology, enabled via /sandbox enable) restrict Copilot’s access to filesystem, network, and system capabilities while running on your machine; cloud sandboxes (copilot --cloud) launch isolated GitHub-hosted Linux environments for stronger security boundaries. Both are included in standard Copilot seats, giving teams agentic workflow security without upgrading their plan.

Copilot CLI: Rubber Duck Agent, Prompt Scheduling, and Voice Input

The GitHub Copilot CLI gets four new capabilities at once: a Rubber Duck agent that critiques plans, designs, implementations, and tests before you ship; /every and /after slash commands for scheduling recurring or delayed prompts (e.g., /every 30m run the frontend tests); hold-space-bar voice dictation with on-device audio processing; and a redesigned experimental terminal UI with tabs for Sessions, Issues, PRs, and Gists. Rubber duck and voice are GA; scheduling and the new UI are via /experimental.

Gemini CLI v0.45.0: Context Management Refactor and A2A Metadata

The June 3 stable release completes a significant refactor of Gemini CLI’s context management system for improved architectural reliability, adds usage metadata exposure to the Agent-to-Agent (A2A) protocol for transparent resource monitoring across multi-agent flows, and resolves critical PTY resize errors and Termux relaunch loops that had been affecting terminal stability. This is likely one of the final major releases before Gemini CLI stops serving free and Google AI Pro/Ultra users on June 18.

OpenCode v1.17.0: WSL Desktop and Faster File Search

OpenCode v1.17.0 (June 10) adds WSL-backed Desktop support and Windows server management, making OpenCode a fully viable option for Windows-native development for the first time. File search is now faster using fff-backed tools, and a revamped sessions and servers UI improves navigation in long-running projects. The follow-on v1.17.4 (June 12) adds workspace-relative directory support for local MCP servers and connector-based authentication with provider credential storage.

Kiro Pro Max: $100/Month Tier and Self-Correcting Subagent Pipelines

Kiro launches a Pro Max tier at $100/month offering 5,000 monthly credits — 2.5x the Pro+ allocation — with premium model access and the full feature set. The same release ships a subagent self-correction pipeline: a reviewer stage can send tasks back to the implementer and loop until the work passes the bar, turning multi-agent pipelines into self-correcting workflows for use cases like code review and refactoring.

📚 Tutorials and Resources

Cursor SDK Updates: Custom Tools, Auto-Review Routing, and Nested Subagents

Cursor’s June 4 SDK update adds significant programmability for teams building on the platform. Custom tools can be registered via local.customTools and exposed to the agent at runtime; auto-review routing lets teams define per-tool permission policies in permissions.json so approved tool calls run without interruption while higher-risk calls escalate; nested subagents are now supported; and JSONL and SQLite custom store options enable persistent agent memory across sessions. Worth reading for any team building internal tooling on top of Cursor’s agent runtime.

Enterprise-Managed Plugins in VS Code: Distributing Agents at Scale

Enterprises can now configure and auto-install Copilot CLI plugins across all licensed VS Code users via settings.json, removing manual setup from developer onboarding. Custom agents, shared MCP configurations, and enterprise-wide hook policies apply automatically — a meaningful governance lever for teams that have been managing plugin sprawl manually. The feature is in public preview and requires VS Code release 1.122 or later.

GitHub Copilot CLI /security-review: On-Demand Security Scanning Before You Push

The new /security-review slash command (experimental, public preview) analyzes local code changes and returns high-confidence findings scored by severity and confidence — covering injection flaws, XSS, insecure data handling, path traversal, and weak cryptography. It runs as an independent AI scan, designed to complement Dependabot and code scanning rather than replace them. The value proposition is speed: findings arrive without leaving the terminal, before code ever hits a PR.

💡 Others

Cursor Design Mode: Select UI Elements, Talk to the Agent, Watch It Edit

Cursor 3.7’s Design Mode inside Canvases lets users select and annotate UI elements directly in-browser — the agent sees the element, surrounding layout, and visual relationships, and can act on voice narration or typed annotations. It closes a workflow gap that has existed since canvas editing launched: the inability to communicate visually about specific components without describing them in text. The companion Context Usage Report breaks down exactly where your tokens go across system prompt, tool definitions, rules, and skills.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Subscribe now

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know that Claude Fable 5 compressed two months of Stripe engineering into a single day, that GitHub’s billing change made the agentic workflows your team relies on dramatically more expensive, and that sub-agents can now spawn sub-agents — five levels deep — across every major tool in your stack. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel.

• The Race Heats Up: xAI enters agentic coding with Grok Build 0.1, Mistral rebrands entirely as Vibe with a unified work-and-code agent, and Codex brings Computer Use to Windows — more direct competition in the same week than any week this quarter.

• The Governance Shift: GitHub now lets enterprise admins assign specific AI models to specific organizations — model governance is no longer an afterthought, it’s a feature Copilot is actively selling to compliance-conscious CIOs.

• The Counter-Narrative: Engineering leaders are quietly capping per-engineer AI budgets as ROI questions surface — the era of unlimited token spend is hitting its first real headwinds.

If you only read one thing this week: Claude Opus 4.8 is the most honest model Anthropic has shipped. It’s approximately four times less likely than Opus 4.7 to let flawed code pass unremarked, scores 69.2% on SWE-bench Pro (up from 64.3%), and introduces Dynamic Workflows in research preview — the ability to orchestrate hundreds of parallel subagents in a single session to complete codebase-scale migrations across hundreds of thousands of lines of code. Fast mode is now three times cheaper than it was on Opus 4.7. The 41-day release cycle from 4.7 to 4.8 signals something: the competition right now is forcing releases faster than the calendar intended. Read more →

Subscribe now

Key Takeaways

• Cursor 3.6’s Auto-review Run Mode makes “run without babysitting” actually safe: A classifier subagent now evaluates every shell, MCP, and fetch call at runtime — allowlisted calls run immediately, sandboxable calls run in the sandbox, and everything else gets escalated to you. The result is fewer interruptions without the anxiety of full yolo mode. Read more →

• xAI enters agentic coding with Grok Build 0.1 — 256K context, $1/$2 per 1M tokens, 100+ tokens/second: Unlike conversational Grok, this model is purpose-built for tool invocation and reasoning chains. It reads diagrams, UI mockups, and error screenshots, integrates with VS Code and JetBrains, and is already compatible with open-source agents like Hermes Agent and OpenClaw via OAuth. Read more →

• Mistral unified its product under one brand called Vibe, and the AI assistant market is converging: Work Mode handles enterprise knowledge across Google Workspace, Slack, GitHub, and more; Code Mode runs remote coding agents and creates pull requests from a web surface or VS Code extension. Every Le Chat account carries over. The race is no longer between chat assistants — it’s between agent operating systems. Read more →

• GitHub’s enterprise Copilot gets model governance at the org level: Admins can now assign specific models to specific organizations instead of applying a single enterprise-wide setting — a change that matters enormously for companies managing different compliance requirements across subsidiaries and teams. Read more →

• Engineering leaders are starting to cap AI token budgets: The Pragmatic Engineer reports that mid-sized and large companies are dampening AI agent spend through per-engineer monthly limits. After two years of uncritical adoption, ROI scrutiny is arriving — and teams that haven’t measured what they’re getting from their AI spend are about to be asked to justify it. Read more →

• Karpathy said “agentic engineering” replaces vibe coding — but Jeff Gothelf noticed it’s just a new name for product management: Writing design specs, supervising agent plans, writing tests, managing permissions — that’s the PM job description. The judgment version of every role is now the job. The administrative version is automatable. Both engineers and PMs need to decide which side of that line they’re on. Read more →

Growing at 20% new subscribers per week.

The stories this week aren’t hard to find. What’s hard is knowing which ones actually matter before your team asks you on Monday.

That’s the only thing Vibe Coding Weekly does: cut through the volume so you arrive at the week with context, not anxiety.

Subscribers also get Change Management in Agentic AI Adoption — the framework for the conversation that always comes after “we should use AI more”: how to actually move an organization that didn’t ask to be moved. Included with every subscription.

Subscribe now

📦 Releases & News

Claude Opus 4.8 Now Available in GitHub Copilot

The same day Anthropic shipped Opus 4.8, it landed in GitHub Copilot for Pro+, Business, and Enterprise subscribers across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, and GitHub’s web interface. Enterprise and Business admins must enable the policy in Copilot settings before their users can access it. A 15X premium request multiplier applies until Usage Based Billing launches June 1 — worth noting for teams already managing Copilot token consumption.

Claude Code v2.1.152–v2.1.158: Opus 4.8, Dynamic Workflows, and Auto-loading Plugins

Five releases in four days. The headlining changes: Opus 4.8 is now the recommended model with /effort xhigh for difficult tasks (v2.1.154, May 28); Dynamic Workflows let you ask Claude to plan work and run tens to hundreds of agents in a single session (same release); /code-review --fix now applies review findings directly to the working tree (v2.1.152, May 27); plugins in .claude/skills/ directories auto-load without needing marketplace publication (v2.1.157, May 29); EnterWorktree can switch between Claude-managed worktrees mid-session; and Auto mode arrives on Bedrock, Vertex, and Foundry for Opus 4.7 and 4.8 (v2.1.158, May 30).

Copilot Memory Gets Deletion Controls, Repo-Level Toggle, and CLI Commands

GitHub gave administrators and users more control over Copilot Memory: a new repository-level off switch lets admins disable memory entirely for a given repo from settings; deletion guidance now points users to the right place to remove a memory and downvotes entries; and three new CLI commands (/memory on, /memory off, /memory show) bring memory management into the terminal. Available to all paid Copilot plans in public preview.

Codex CLI 0.134.0 and 0.135.0: History Search, Vim Mode, and Computer Use on Windows

Two releases in rapid succession. 0.134.0 (May 26) adds local conversation history search with case-insensitive content matching, makes --profile the canonical selector across all CLI flows, and lets read-only MCP tools run concurrently via readOnlyHint. 0.135.0 (May 28) adds richer codex doctor diagnostics and named permission profiles. The May 29 app update brings the most notable change: Computer Use now works on Windows — Codex can see, click, and type in Windows desktop apps — plus remote control of Codex sessions from iOS, Android, or Mac.

Gemini CLI v0.44.0: Unified Auto Mode and Sublime/Emacs Support

The May 27 release merges all specialized Auto modes into a single unified mode, simplifying configuration for teams running mixed workflows. Native support for Sublime Text and Emacs Client arrives as first-class integrations, and new agent-tui and tui-tester skills enable programmatic testing and automation of terminal UI applications — useful for anyone building or testing TUI-based developer tooling.

📚 Tutorials and Resources

Dynamic Workflows Explained — What They Are and How to Use Them

Appwrite’s breakdown of Opus 4.8’s most consequential new feature: Dynamic Workflows let Claude plan and execute work across tens to hundreds of parallel subagents in a single session. The piece explains the practical use cases — codebase migrations across hundreds of thousands of lines, full-project refactors that would previously require manual task decomposition — and clarifies that this ships as a research preview available today in Claude Code. A useful orientation if you’re trying to understand when and why to reach for this versus a standard agentic flow.

Next week, the stack keeps moving. So does this newsletter. Fall behind one week, and you’ll spend the next three catching up.

Every week, a new model drops. A new agent framework ships. A new “this changes everything” thread goes viral. And you still have actual code to write.

Every Monday, you open your inbox and already know what matters. You’ve skipped three viral threads that turned out to be nothing. You know that Opus 4.8 can run hundreds of parallel subagents on your codebase in a single session, that engineering leaders are quietly starting to cap AI token spend, and that every AI assistant on the market is converging toward the same thing — an agent operating system — whether they admit it yet or not. You didn’t spend your weekend reading to know this. We did.

That’s what Vibe Coding Weekly is. For developers, architects, tech leads, and everyone building or managing software in the age of AI.

Clean code and positive vibes,
Angel.