How Anthropic's Claude Code team runs review when agents write the code

Addy Osmani and colleagues on the Claude Code team wrote up how they review code now that Claude writes about 90% of Claude Code itself. The goal they optimize for is blunt: the least human review time that does not raise the revert rate or cause more incidents. Review has stopped being a gate at the end and runs continuously instead, closer to a spellcheck that works while you type.

The numbers they share are the reason to read it. Merged pull requests to Claude Code more than doubled between June and September while the revert rate fell by more than 40%. Across Anthropic's main code repository, daily changes went from about 2,500 in May to over 30,000 in September. After they launched an AI reviewer, the share of internal PRs getting substantive review comments rose from 16% to 54%, and the share that passed after a single round of fixes went from 65% to 78%.

The process is layered. Claude runs tests and a fresh-context review before a PR opens, parallel agents hunt for bugs and vote on each finding, security findings can block a merge, and a classifier assigns each change a risk tier that decides how many humans must approve. Agent approvals never count as a human approval. One safeguard exists because an early reviewer once ran a PR through 50 or more rounds of agent-versus-agent fixes overnight: after three rounds, a human reconsiders the approach. The full writeup includes their CLAUDE.md and REVIEW.md conventions.

Why it matters

If your team is starting to merge agent-written code, the hard part is review, not generation. Anthropic's setup is a concrete template: risk tiers that decide human involvement, security gates that block merges, and old incidents turned into a daily eval for the reviewer itself.