How the Claude Code team does code review
Anthropic describes a review process that moved routine checks toward machines as pull-request volume grew, while retaining human judgment for high-risk decisions.
TL;DR
- Anthropic says merged pull requests to the Claude Code repository more than doubled from June to September while the share later reverted fell by more than 40%.
- The company describes layered AI review, automated checks, risk-based human approval, feature flags and rapid rollback.
- These are company-reported internal results; the post does not provide a controlled comparison establishing which changes caused the reported outcomes.
Anthropic engineers published an account of how the Claude Code team changed its review process as AI-generated code increased. The company says merged pull requests to Claude Code’s repository more than doubled between June and September, while the share of changes later reverted fell by more than 40%. [1]
The process shifts much review work to automated checks: Claude reviews pull requests before a person does, separate security checks can block a merge, and test, lint and build results are required. Human approval is adjusted to the risk of the change, with flags and staged releases used to limit impact. [1] [2]
The authors report that their internal monorepo’s daily changes rose from about 2,500 in May to more than 30,000 in September, and say automated reviewers are continually evaluated against past incidents and reverts. Hacker News readers discussed the approach and what other teams use for review. [1] [2]
The figures and outcomes are Anthropic’s own internal measurements, and the post does not isolate the effects of AI review from other process changes. Whether similar results hold for teams with different codebases, risk levels and staffing remains unknown. [1]
Why it matters
The account offers a concrete example of a software team shifting human review toward risk, architecture and accountability as code volume rises. The reported gains are internal company figures rather than an independent evaluation.
Editor's note
Attributes process details and performance figures to Anthropic; notes that the post does not establish causation or generalizability.