How the Claude Code team does code review — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic's Claude Code team restructured review after acknowledging Claude writes about 90% of the codebase, optimizing for the minimum human review time that doesn't raise reverts or incidents.
- Merged pull requests to Claude Code more than doubled from June to September while the share requiring reverts fell by more than 40%, according to internal data the team shared.
- Anthropic's monorepo saw changes landing per day jump from about 2,500 in May to more than 30,000 in September, the volume that pushed the team to move most code reading before a PR is even opened.
- A Claude automated reviewer reads every PR diff against the full repo and stated intent before any person looks; when launched as a research preview, the share of internal PRs getting substantive comments rose from 16% to 54%.
- A risk-tier classifier determines how much human review each change needs — some paths always require a second person, and agent approvals never count toward the human slot.
- Past incidents are now treated as automated evals: engineers confirmed each bug and where it lived, and the Claude reviewer runs against that set daily to catch the same failure modes.
- New behavior usually ships behind a flag to Anthropic employees first, with a merge queue that re-runs every check against main to keep two passing PRs from combining into a broken trunk.
Why it matters: The shift reframes code review from line-by-line gatekeeping to risk-routed exception handling: with 90% of code AI-written, humans no longer read most diffs, they classify risk and approve or catch what the agent stack missed. The over-40% drop in reverts alongside doubled PR volume is the strongest evidence yet that automated review can hold up at Anthropic's actual shipping cadence, not just in demos.
Ask SkimNews
