How the Claude Code team does code review
Addy Osmani, Morgan Westlee Lunt and Qing Wang · 15 October 2026
Read the originalThe Claude Code team at Anthropic sets how many people approve each change from what the change touches, and doubled its merged changes while reverts fell.
Claude Code is Anthropic's AI coding agent, and Claude writes about 90 percent of its code. Addy Osmani, Morgan Westlee Lunt and Qing Wang describe how the team reviews that volume. Machines do most of the reading, and the number of human approvers depends on how risky each change is.
- Each change passes AI review, a security review that can block the merge, and automated tests before a person is asked to look.
- A classifier sets the tier from the change itself: none for documentation, one person for ordinary product code, two for the riskiest paths.
- An AI agent's approval never fills a slot that needs a person, and whoever directed the agent counts as one of the reviewers.
- From June to September, merged pull requests more than doubled while the share that had to be reverted fell by more than 40 percent.
A team can write down its approval tiers by risk and let a tool assign them, provided rollback is fast. The account comes from Anthropic, which sells the review product it describes.
Write down approval tiers by risk, set from what a change touches rather than who wrote it, and require two people only for the riskiest paths.
After three rounds of AI review and fixes on one pull request, have the AI reviewer stop and ask the engineer who wrote it to rethink the approach.
Derived by Working Surface from the article; more in the Playbook. Source line: An agent's approval never counts as a person's, either.
18 October 2026: as AI agents make code and prototypes cheap, teams are writing down which work needs a human check and which ideas have earned more effort.