How the Claude Code team does code review

Claude Code’s engineering team has shifted from a traditional gate‑at‑the‑end code review model to a continuous, automated review process that functions like spell‑check while developers write code. The change was driven by the rapid increase in code changes—from about 2,500 daily in May to over 30,000 in September—making manual line‑by‑line review impractical. The new workflow begins before a pull request (PR) is opened: the agent runs tests, builds, or visual comparisons and presents the output to the author. A subagent then reviews the diff against the entire repository and the PR’s stated intent, flagging potential issues. Findings are verified by multiple agents and only posted if a consensus is reached, filtering out style or diagnostic comments. Each finding is tagged by severity and must be addressed before the PR can merge. Security reviews run in parallel, and any open findings block merging. The system classifies each change into risk tiers—Quick, Standard, or Deep—determining whether a human reviewer is required. Only trusted authors can bypass the human review path; otherwise, a second human approval is mandatory. The process tracks human approvals, comments, and wait times, using revert and incident rates as guardrails to ensure that human effort is not merely shifted. The goal is for most PRs in the Claude Code repositories to merge with at most one human in the loop, maintaining current revert and incident rates while reducing overall human review time. This approach illustrates how AI can handle high‑volume code production while preserving critical human oversight, a strategy that may influence broader AI industry practices.