How the Claude Code team reviews code at agent speed

Claude writes about 90% of Claude Code's own code. From June to September, merged PRs to the codebase more than doubled while the share reverted fell by more than 40%. The team moved most of the reading to machines: verification before a PR exists, AI review on every PR with findings voted on by other agents, blocking security review and CI, human approvals scaled by risk tier, then merge queue, flags, staged releases and fast rollback. Why it matters: It reframes review around one budget, human time and attention, with revert and incident rates as the guardrail. A classifier sets the risk tier from the change, not the author: low-risk changes need no human approval once checks are green, and sensitive paths need one or two people. An agent's approval never counts as a human's. Past incidents became a daily eval for the reviewer, and a three-round cap stops agent authors and reviewers from looping. One caveat is that these are Anthropic's own internal numbers on its own product.

Try this: Write down three risk tiers for one repo, set from what the change touches rather than who wrote it, and add "Testing", "Risk" and "How to revert" to its PR template.