How the Claude Code team does code review
How the Claude Code team does code review
The Claude Code team describes a review path that begins before a pull request: agents run tests and inspect the diff with fresh context. On each PR, parallel agents search for bugs, separate agents verify their findings, and security and ordinary CI checks run too. Findings must be fixed or closed with a reason. The team limits repetitive review rounds after one PR cycled more than fifty times; an author is asked to reconsider the approach after three. A merge queue checks the combined change, then flags, internal use, staged channels and on-call rollback cover mistakes that get through.
Human approval depends on the change’s risk. Trusted authors can land certain low-risk changes without a second person; sensitive paths and higher-risk changes need one or two human approvals. The person directing an agent counts as a reviewer, but an agent cannot fill a human approval slot. The authors reserve people for deciding whether to build something, setting architecture, checking what test evidence misses and owning repairs. They say Claude Code’s merged PRs more than doubled from June to September while the share reverted fell more than 40%, and report fewer machine-review rounds after changing the process.
These are first-party before-and-after observations, not proof that automated review caused fewer problems. The article does not publish active human review and repair minutes, absolute revert or incident counts, false-positive costs or independent customer acceptance; closing a finding with a reason need not mean the behavior was fixed. There is not yet an automatic circuit breaker tightening approvals when quality worsens. This is the Claude Code repository with a merge queue, not Anthropic’s separate monorepo, whose integration checks use a different order.