# Review agent-written code in layers the way the Claude Code team does

PaperChase agent use case (free feed entry), version 1.0. Source: [Addy Osmani, Morgan Westlee Lunt and Qing Wang (Claude Code team, Anthropic)](https://claude.dev/blog/solving-code-review-at-anthropic/), 2026-10-15.

## What they did

Addy Osmani, Morgan Westlee Lunt and Qing Wang laid out a stack of checks: the agent must show test or build output before a PR, a fresh-context subagent reviews the diff, several review agents look for bugs and a second group votes before a finding is posted, security findings block merge until fixed or dismissed with a reason, and a classifier sets a risk tier that decides whether zero, one or two humans approve. Merge queue, flags, staged release and fast rollback sit underneath. Past incidents became a daily eval for the reviewer.

## Worth knowing

Review has to converge: one PR went through more than fifty agent rounds overnight, so they now cap rounds at three, re-review only new diffs, and suppress repeat findings. An agent's approval never fills a required human slot, and whoever prompted a change owns it. Internal numbers are Anthropic's own. The post is dated Oct 15 with no time; the earliest share we found on X was Oct 16 at 01:49 UTC.

## Workflow template

```text
Adoption steps the post lists, in our words: keep a short CLAUDE.md with checks the agent can run; run /code-review and /security-review locally before a PR; add a REVIEW.md and move repeated review comments into it; require CI and add Testing, Risk and How to revert to the PR template; turn on an AI reviewer for one repo first; require every review thread resolved before merge; write risk tiers based on the change, not the author; cap review rounds; turn incidents and reverts into an eval set.
```

## Steps

1. **Read what they did.** Skim the source post linked below, then the summary on this page.
2. **Fill in the template.** Copy the prompt and replace every [bracketed] value with your own project details.
3. **Run it once and check.** Run it on a small task first and confirm the result yourself before you trust it on real work.

## Credit

Workflow reported by Addy Osmani, Morgan Westlee Lunt and Qing Wang (Claude Code team, Anthropic): https://claude.dev/blog/solving-code-review-at-anthropic/
Related: https://news.ycombinator.com/item?id=50122951
Related: https://x.com/bcherny/status/2111158313463525518

Summary and template written by PaperChaseWebb in our own words. Questions: mgmt@paperchasewebb.com