Agent Patch Weekly — Issue #002

Patch notes for the agent meta · Week of Oct 21, 2026

GG and welcome back to patch day. This one covers Oct 14 to 21: Claude Code's Code Review went GA, GitHub deprecated a stack of Copilot models, Google open-sourced a harness that tunes itself, and a solo dev posted the most honest game launch numbers we've seen. Plus one workflow we actually ran. Every item links to its source.

🟢 BUFFS

Claude Code: Code Review is generally available (Oct 15) Anthropic's automatic PR reviewer is out of research preview for Team and Enterprise plans. Anthropic says most reviews now finish about 3x faster (~20 minutes down to 6–10), and admins pick Quick, Standard (the default) or Deep for each repo. On any plan, /code-review runs a review inside your own Claude Code session. → Anthropic announcement

Claude Code v2.1.304 and v2.1.305 (Oct 19–20) New /auto-mode-setup reads your project and recent sessions to teach auto mode about your environment. v2.1.305 fixes every request failing on Vertex, Bedrock and other providers when a gateway rejects tool search, a regression from 2.1.304. If 2.1.304 broke your setup, update. → Claude Code changelog

Cursor: canvases from cloud agents (Oct 16) Cloud agents can now build interactive canvases and share them by link. The Environments API is in beta, cloud environments support up to 300 repositories, and large ones load 5x faster. → Cursor changelog

T3 Code leaves alpha, then ships a 1,400-change stable (Oct 15 and Oct 20) T3 Code dropped the Alpha label with 0.1.0 Beta and put Orchestrator V2 on the stable channel. On Oct 20, Theo shipped a stable release with over 1,400 changes, including perf wins and idle Claude threads that compact early so the cache stays warm (beta). Before that, Julius from the T3 Code team said his agent merged almost 400 web-perf PRs in 30 hours: 86% fewer renders when another thread streams, up to 35% less JS CPU and a ~10% smaller heap. Theo also says over 60% of T3 Code users who used Codex in the last 2 weeks also used Claude Code. If you run more than one agent, keep one rules file they all read (our guide). → 0.1.0 Beta · stable release · 400 PRs · 60% stat

🔴 NERFS (and heads-ups)

GitHub Copilot deprecated six models on Oct 19, and three more go Nov 16 Gone as of Oct 19: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5. On Nov 16: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra (Opus 4.8 stays for individual subscribers on annual plans). GitHub's suggested replacements include GPT-6.1 Sol, GPT-6 Luna, Gemini 3.8 Flash, Grok 4.7 and Claude Opus 5.5. If a script or workflow pins a model name, check it now. → Oct 19 deprecations · Nov 16 deprecations

Replit makes OpenAI its default (Oct 15) At TechCrunch Disrupt, Replit CEO Amjad Masad said Replit is no longer working with Anthropic. He later clarified that Replit stays model agnostic and users can still pick Anthropic models, but OpenAI is now the preferred partner powering Replit Agent. TechCrunch notes that most customers stick with the default, and that insiders disputed the on-stage claim. Know which model your agent runs by default. → TechCrunch · Masad's clarification

The frontier edge is shrinking (Oct 18–19) The Information reports that a 10-point intelligence advantage now earns just 20% more developer traffic, down from nearly 69% in late 2025. Cognition's Russell Kaplan says that's consistent with their internal data: many software engineering tasks are now "intelligence saturated." In his words: "Once any model can do the task well, cost and speed are all that matter." Our take: for routine tasks, try the cheaper, faster model first. → The Information · Russell Kaplan

Claude Code heads-ups: resume costs and Claude Design's end date v2.1.304 (Oct 19) warns Pro and Max users when resuming an inactive conversation would use a noticeable share of the 5-hour usage limit. You can resume anyway, or start fresh and Claude can look up the old conversation. v2.1.302 (Oct 16) added a notice that standalone Claude Design closes by Dec 14, 2026. → Claude Code changelog

🎁 NEW LOOT (models and tools)

🛠️ WORKFLOW OF THE WEEK: The Proof-of-Done Loop

The problem: your agent says "done." Did it run the tests? Did it change a file after they passed? The fix: copy the Claude Code team's first review rule. Claude writes about 90% of Claude Code's code, and from June to September merged PRs more than doubled while the revert rate fell by more than 40%. Their first layer: give the agent something it can run and make it show the output instead of saying it's done (their write-up). This loop adds a receipt you can check.

  1. Install the free Checkpoint Loop, which gives you scripts/agent-check.sh. Save the script below as scripts/proof.sh, run chmod +x scripts/proof.sh, and add .agent/ to .gitignore.
  2. Add this to your AGENTS.md:
    ## Proof of done
- Never say "done" without proof. Run ./scripts/proof.sh and paste .agent/proof.txt in your final message.
- If the result is RED, fix it and run it again. Stop after 3 failed fixes and tell me what's blocking you.
- If you change any file after the proof, run it again.
  3. Before you accept the work, run ./scripts/proof.sh --verify. PROOF OK means the check was green and no file has changed since. STALE PROOF means something changed after the check ran.
  4. Then get fresh eyes: run /code-review in Claude Code, or have a fresh-context subagent review the diff. Cap it at 3 rounds, like the Claude Code team does.
#!/usr/bin/env bash
# proof.sh: run the check and save a receipt your agent must show before it says "done".
#   ./scripts/proof.sh           run the check, write .agent/proof.txt
#   ./scripts/proof.sh --verify  pass only if the receipt is GREEN and matches the code as it is now
set -uo pipefail
cd "$(git rev-parse --show-toplevel)" || exit 1
mkdir -p .agent
SHA="sha256sum"; command -v sha256sum >/dev/null || SHA="shasum -a 256"  # macOS fallback
state() {  # fingerprint of every tracked and new file's contents (ignores .agent/), so committing doesn't change it
  git ls-files -co --exclude-standard --deduplicate -z -- . ':!.agent' | sort -z | xargs -0 -r $SHA 2>&1 | $SHA | cut -c1-16
}
if [[ "${1:-}" == "--verify" ]]; then
  [[ -f .agent/proof.txt ]] || { echo "✖ NO PROOF: run ./scripts/proof.sh"; exit 1; }
  grep -qx 'result: GREEN' .agent/proof.txt || { echo "✖ PROOF IS RED"; exit 1; }
  grep -qx "state: $(state)" .agent/proof.txt || { echo "✖ STALE PROOF: the code changed after the check ran"; exit 1; }
  echo "✔ PROOF OK: green, and it matches the current code"; exit 0
fi
out=$(./scripts/agent-check.sh 2>&1); code=$?
{ echo "date: $(date -u +%FT%TZ)"; echo "state: $(state)"
  echo "result: $([[ $code -eq 0 ]] && echo GREEN || echo RED)"
  echo "--- files changed"; git diff HEAD --stat -- . ':!.agent'; git ls-files -o --exclude-standard -- . ':!.agent'
  echo "--- check output (last 40 lines)"; echo "$out" | tail -n 40; } > .agent/proof.txt
cat .agent/proof.txt; exit $code

What we actually tested: 10 cases on a sample npm repo and a sample Python repo (Linux, git 2.47, Node 20, Python 3.13). A green proof verified OK, and committing afterwards still verified OK. Editing a file, adding a file or deleting a file after the proof all came back STALE. A broken function came back RED, and so did a fake API key in the diff. Fixing and re-running went back to OK, and a missing proof was rejected. We have not tested it on macOS, and we haven't measured how reliably each agent follows the rule. It's a tripwire for honest mistakes, not a lock: an agent that edits the receipt could fool it, so rerun ./scripts/proof.sh yourself on anything that matters.

🎮 GAME DEV CORNER

Bolt Pit: honest numbers from a Claude Code game launch (Oct 19) Frederik (@froessell) built the browser game Bolt Pit with Claude Code. It's been live on CrazyGames since Oct 15: 664 plays from 628 players and $0 earned so far (it's in a no-ads Basic Launch), with 3:00 average playtime, 31.41% gameplay conversion and 0.70% day-one retention. His agent traced each number to the code. Bots clear all six waves in 115 to 125 seconds, so one run is about all the content there is. About 14 MB loads before play, including a 5 MB boss model you don't meet until wave 6. And nothing gives players a reason to come back tomorrow. His fix plan has one phase per number, conversion first, one commit each. It's the best launch post-mortem we read this week. → Frederik's breakdown on X

Tim Sweeney: "a new golden age of gaming" (Oct 17) Epic's CEO: "The games we've long wanted, which didn't exist because they'd cost several times more to make than they'd earn back, will soon become possible." Cheap to build still isn't the same as fun to play, and Bolt Pit's retention numbers are the reminder. → Tim Sweeney on X

UEFN: 2–3 months down to 2 weeks (Oct 18) UEFN creator Tora: "One game used to take me 2-3 months to finish in UEFN, now it takes no more than 2 weeks." He credits AI and MCP. → Tora on X

Roblox: "All your favorite Roblox games use AI now" (Oct 18) KreekCraft says he can name maybe 3 or 4 games he knows for sure don't. Nikilis, the creator of Murder Mystery 2, added the nuance: MM2 uses AI for code only, and none of its art, concepts, animations or UI. AI writes the code; taste is still the moat. → KreekCraft · Nikilis

An Xbox manager's AI thread, then a deleted account (Oct 18–19) Brandon Stirrup, a senior business strategy manager at Xbox, posted that he'd "vibe coded a bad but improvable game" with about three hours of active work, and that "we use AI, just as all companies do in 2026." After a flood of replies he deleted his X account; a source at Xbox told PC Gamer that was because of the volume of replies. It was his own thread, and we found no official Xbox statement about it. → PC Gamer · Kotaku

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

Bolt Pit's numbers show where an AI-built game can fall short: playtime and replay, not build speed. Pack #001 has seven agent templates to take a game from idea to playable, fun and shippable: spec, AGENTS.md for game projects, playtest agent, "Is It Fun?" review (loop, hook and replay pull), juice pass, bug triage and ship checklist. → Get the pack ($19) · See what's inside

📦 Free: Template #001, The Checkpoint Loop. This week's workflow builds on it. → Download free

Want to sponsor an issue? Hit reply. Hit reply and tell us about the worst "done" your agent ever gave you. We read everything. Agent Patch Weekly is researched and drafted with AI agents and reviewed before sending. Every link was checked on Oct 21, 2026.

← Previous: Issue #001 · All issues

Get the patch notes free →

Free, weekly, one-click unsubscribe.