Anthropic’s Claude Code team now runs code review continuously rather than as a final gate. Most review happens before a pull request exists, with agents verifying work and reviewing diffs. Every pull request gets an automated Claude review before human eyes see it. Security review can block a merge until findings are fixed or cleared. Automated checks like lint, type checks and tests must pass before review. Human approval scales by risk: none for low, one for medium, two for high. A merge queue re-runs checks against main to prevent combined breakage. New behavior ships behind flags to employees first, then to the latest channel. Rollback is fast, with flags and on-call rotation catching issues quickly. The team measures human time via approvals, comments and wait time, with revert and incident rates as guardrails. Review agents run nightly over existing code to catch old bugs. Noise is measured by tracking how authors respond to comments. A round cap of three stops infinite agent-author loops.
What changed
Manual line-by-line code review was the default gate before merging.
What it unlocks
Merging most pull requests with at most one human in the loop without raising reverts or incidents.
- merged PRs more than doubled Jun–Sep 2026
- reverts fell by over 40%
- 2,500 to 30,000 changes per day in Anthropic monorepo
- 59% of repeated flags were pre-existing code
What you need to act on it
- Claude Code Team or Enterprise plan for automated code review
- Trusted author status for lighter human review paths
- Automated review and security checks must be green for low-risk changes
Sources