How the Claude Code team reviews code: ~90% Claude-written, merged PRs more than doubled June–Sept while revert share fell over 40%

Anthropic's Claude Code team (Oct 15 post by Addy Osmani, Morgan Westlee Lunt and Qing Wang) says Claude writes about 90% of Claude Code's own code, and explains how they review it at that volume. Review moved from a final gate to continuous layers: pre-PR verification (tests, build, lint) plus a fresh-context subagent review; a Claude review on every PR where parallel bug-hunting agents are filtered by voting verifiers and tagged Important/Nit/Pre-existing; a separate security review whose open findings block merge; required CI; a risk classifier that sets 0, 1 or 2 human approvals based on the change, with sensitive paths and CI workflows always needing a person and agent approvals never counting as human ones; then a merge queue, flags, internal dogfooding, latest-before-stable releases and fast rollback. Internal metrics (no raw data published): merged PRs more than doubled June–September while the revert share fell 40%+; after a three-round cap, diff-only re-review and no repeated findings, one-round pass rose 65%→78% and mean reviews to pass fell 2.08→1.32. Past incidents and reverts form a daily eval for the reviewer. The team admits it lacks circuit breakers that tighten approval when revert or incident rates rise.