Claude writes 90% of Claude Code, so Anthropic rebuilt code review

Anthropic published a write-up on 15 October explaining how its Claude Code team reviews code now that Claude writes about 90% of it. From June to September, merged pull requests to the Claude Code codebase more than doubled, while the share that had to be reverted fell by more than 40%. Across Anthropic’s wider monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September.

The core mechanism is a risk tier assigned to each change, not each author. Low-risk changes, docs, copy, anything behind a flag, need no human approval once automated checks pass. Medium-risk changes need one person, high-risk changes need two. Before a pull request even exists, Claude runs tests and a linter and a second agent reviews the diff in fresh context. Every PR also gets a round of AI review where several agents hunt for bugs and a separate group votes on whether each finding is real before it gets posted.

One anecdote stands out: a single CLI pull request went through more than fifty rounds of agent review and fixes overnight before a human stepped in. Investigating afterwards, the team found the reviewer was mostly re-flagging issues that had already been there since its first pass. That led to a three-round cap and diff-only re-review, which dropped the mean number of review rounds needed from 2.08 to 1.32.

Why it matters: if your team is already shipping AI-authored pull requests, this is a working model for where human review time still earns its keep, risk-tiered approval rather than reviewing everything at the same depth, plus a concrete lesson about agents re-litigating old findings.

The caveat: Anthropic says it doesn’t yet have circuit breakers that tighten approval automatically if revert or incident rates start to climb, and admits review’s second job, building shared context across a team, still has no formal process of its own.