How the Claude Code team reviews code at agent volume
Summary¶
Anthropic's Claude Code team moved most code reading to machines: parallel AI reviewers whose findings are verified by agent voting, a blocking security review, and human approval tiered by a classifier that reads the change (none, one or two people). Convergence rules (a 3-round cap, diff-only re-review, no repeat findings) cut mean reviews-to-pass from 2.08 to 1.32, while merged PRs doubled and reverts fell by more than 40%.
Key Details¶
- Risk tiers are set from the change, not the author; CI workflow files and owner-listed sensitive paths always need a person, and a bot approval never fills a human slot.
- An approval covers one commit; an agent push or a hand-resolved merge requires re-approval.
- About 59% of repeated reviewer flags were pre-existing code; pass after one round rose from 65% to 78% and PRs needing 5+ rounds fell from 8.9% to 1.5%.
- Past incidents, each confirmed by an expert, form a daily reviewer eval; noise is measured from how authors respond to each comment.
- Code Review for Claude Code is generally available on Team and Enterprise plans;
/code-reviewand/security-reviewrun locally.