Review AI-written code with a layered agent review process
Anthropic's Claude Code team describes how it changed its code review process so Claude can write most of Claude Code's code.
Done withClaude CodeClaude
- What they did
- Claude verifies its own work by running tests, builds or screenshots, and a fresh-context subagent reviews the diff. Every PR then gets parallel bug-hunting agents, with other agents voting on each finding, plus a blocking security review and CI. A classifier sets human approval to none, one or two people by risk. Merge queue, flags and gradual release follow.
- How it went
- From June to September, merged PRs more than doubled and the share reverted fell by over 40 percent. Convergence changes cut mean reviews to pass from 2.08 to 1.32. Circuit breakers that tighten approval don't exist yet.
- Worth knowing
- Agent author and agent reviewer can loop: one PR went over fifty overnight rounds, so cap re-review rounds (they use three) and re-review only the new diff.
Try it yourself with Claude Code
Review the pull request on branch [branch name] in this repo in layers. First check it against the intent described in [PR description or ticket link], then look for bugs, missing tests and security issues, and finally check style against our conventions in [file or doc]. Finish with a list of issues sorted by severity and a verdict on whether it is safe to merge; do not change any code or push anything.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on HN · Oct 16, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Audit subscriptions and email companies to cancel them
Claude Code pulled transactions across credit cards, flagged recurring charges, and drafted cancellation emails, recovering $700.
Investigate market moves and alerts with an agent on your charts
OpenChart lets Claude or Codex annotate chart setups, write indicators, and investigate price alerts, saving research to a local workspace.
Review Linux kernel patches
Google engineers launched Sashiko, an agentic AI code reviewer for the Linux kernel.
Custom agent
Turn customer feedback into tested pull requests
In OpenAI's own launch example, a Dot watches feedback for recurring requests, scopes small fixes, builds and tests them, and returns PRs with demo videos.
Build a book donation app and demo video in a buildathon
A user used Replit Agent to build the GiveABook app and its demo video within a 24-hour buildathon.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.