Agent Patch Weekly — Issue #001

Patch notes for the agent meta · Week of Oct 16, 2026

GG and welcome to patch day. Every week we read the changelogs so you don't have to. We sort them into buffs, nerfs and new loot, then hand you one workflow we actually ran. Every item links to its source.

🟢 BUFFS

Cursor: Origin App Marketplace (Oct 13) Origin now plugs into 20+ apps for deploys, CI, code review, security, monitoring and project management. Partners include Cloudflare, Vercel, Linear, Sentry, CodeRabbit and Semgrep. The marketplace is also open to public submissions. → Cursor changelog

Cursor: remote control for local agents (Oct 6) You can see and reply to the agents running on your computer from the Cursor iOS app. It's on by default except for Enterprise orgs. Your computer has to stay on and online. → Cursor changelog

Codex CLI 0.163.0 (Oct 15) You can steer Codex mid-response by sending a follow-up. It also gets a searchable /config, a configurable leader key (default Ctrl+X), and parallel read-only tools. Daybreak is on by default for eligible accounts. → openai/codex releases

Claude Code: a big October so far v2.1.287 added Claude Mods (plugins can modify deeper behavior). v2.1.293 made Claude Haiku 5.5 the default Haiku model on the Anthropic API, with 1M context. v2.1.295 added onFailure: "block" for hooks. v2.1.292 added claude plugin install --marketplace and an effort parameter for sub-agents. v2.1.302 shipped Oct 16 with more auto-mode controls. → Claude Code October release notes (Havoptic)

🔴 NERFS (and heads-ups)

Gemini 4 Argon's price doubles after Nov 30 The launch price is $2 / $10 per 1M input/output tokens. After the introductory period ends on Nov 30, 2026, it goes to $4 / $20. Plan your long agent runs now. → Google blog

ChatGPT Pro 200: lower usage allowance, same $200/mo The tracker Aftergradient reports a lower usage allowance on the current Pro 200 offer, with impact on Codex (posted Sep 29). Check your plan page before you schedule heavy Codex sessions. → Aftergradient updates

Claude Code idle compaction is on Since v2.1.299, conversations over 200K tokens get compacted after about 54 minutes idle. If you want the full context back after lunch, set "idleCompaction": false. → Havoptic

Codex dropped its patched zsh 0.163.0 retires the bundled patched zsh. Configs that use the retired shell flags get a migration notice. → openai/codex releases

🎁 NEW LOOT (models and tools)

🛠️ WORKFLOW OF THE WEEK: The Checkpoint Loop

The problem: an agent makes 14 changes in a row, and somewhere around change 6 it broke the build. The fix: make the agent prove each step is green before it moves on.

  1. Drop AGENTS.md and scripts/agent-check.sh into your repo. Claude Code v2.1.277+ reads AGENTS.md on its own when there's no CLAUDE.md (guide). On older versions: ln -s AGENTS.md CLAUDE.md.
  2. Start every session with:
    Read AGENTS.md and follow the Checkpoint Loop exactly.
Task: <feature/bug>
Done means: <observable result>
Start with step 1 (Plan) and wait for my "go".
  3. The agent works on an agent/<task> branch in small steps, and runs ./scripts/agent-check.sh after each one. It commits only on GREEN and stops after 3 failed fixes.

What we actually tested: we ran the check script on a sample Python repo and a sample npm repo. It returned green on passing tests, red on a broken function, and red on a fake API key in the diff. Full results are in the README. We have not benchmarked how closely each agent follows the rules; that varies by tool.

📦 Free download: Template #001, The Checkpoint Loop → get it on the templates page

🎮 GAME DEV CORNER

Google and Unity launch Playground (Oct 7). Playground is an experimental platform where users 18+ make playable games from text prompts, no coding required. Later this year it gets Unity Spark, an expanded creation experience with access to the Unity runtime. → Unity press release

Roblox shares how it grades AI-made games (Oct 8). Build, Roblox's prompt-to-game tool, scores games on 4 signals: spec quality, static code/design, playtests driven by real player-like input, and visual quality. Roblox says about 9,000 Build games have been published since the New Zealand alpha. The best detail: one obby scored 98/100 on engineering but 31/100 on design. It worked, but it wasn't fun. Steal this rubric for your own agent-made games. → Roblox newsroom

Unity's official agent plugin lands in Grok Build (Oct 1). Grok Build gets the same 30+ Unity-written skills as Claude Code and Codex, for Unity 6 and above. Install: grok plugin install Unity-Technologies/unity-agent-plugin --trust. → Unity blog

Summer Engine 4.0 (Oct 16): "the game engine for AI agents." Any MCP-capable agent can build, generate assets, publish and read analytics. Publishing, live changes, deletes and contacting people all ask for your approval first. Connect with claude mcp add --transport http summer https://mcp.summerengine.com/mcp. → Summer Engine blog

📦 Pack #001: Vibe Coder Game Jam Starter ($19)

Seven agent templates to take an AI-built game from idea to playable, fun and shippable: spec, AGENTS.md, playtest, fun review, juice pass, bug triage and ship checklist. → Get the pack ($19)

Hit reply and tell us which agent you main. We read everything. Agent Patch Weekly is researched and drafted with AI agents and reviewed before sending. Every link was checked on Oct 16, 2026.

Next: Issue #002 → · All issues

Get the patch notes free →

Free, weekly, one-click unsubscribe.