Claude Code team doubles merges and cuts reverts by 40% with AI review system

Details

The Anthropic Claude Code team achieved a result of more than doubling the number of merged PRs while reducing the revert rate by over 40%.
They applied a six-layer inspection structure focused on 'what could go wrong,' using AI reviews and automated checks to reduce the burden of human reading and focus on high-level review.
After introducing AI reviews, the rate of receiving substantive review comments increased from 16% to 54%, and 84% of PRs with over 1,000 lines received an average of 7.5 points of feedback.
To prevent infinite loops between agents, convergence measures were introduced to reconsider the approach after the third iteration and to re-review only the diff after the initial iterations.
With the application of convergence measures, the first-pass pass rate improved from 65% to 78%, and the average number of review iterations decreased from 2.08 to 1.32.
Humans are responsible only for five core roles: determining direction, reviewing risky code and architecture, verifying evidence, and taking responsibility for results, while code understanding is resolved by requesting Claude to generate documentation.
The Code Review product uses a token-based billing model, with an average cost of $15–$25 per review.
Only GitHub repositories are supported, and organizations with Zero Data Retention or HIPAA compliance requirements are recommended to use alternative tools.

Key point