The post reports that merged pull requests to Claude Code repositories more than doubled from June to September, while the share later reverted fell by more than 40%. The team says it moved much of review earlier and delegated routine checks to automated layers rather than removing review.
Before a pull request, Claude runs tests, a build and a linter, and a second agent reviews the diff in a fresh context. Pull requests then pass required CI, parallel Claude code-review agents, a separate security review and risk-based approval. A classifier sets the human approval tier: none, one or two, with some paths always requiring a person. After merge, feature flags, gradual rollout and rollback provide recovery controls.
Anthropic measures human approvals per pull request, comments and time waiting for a person. Reverts and incidents are guardrails: reducing review time is not a gain if failure-response work rises. The post describes the Claude Code team’s process; it does not establish that the same results will generalize to other organizations.