According to reports, Anthropic stated that Claude has written approximately 90% of the code for Claude Code. From June to September this year, the number of merged code pull requests (PRs) increased by over 100%, while the rate of reversions dropped by more than 40%. To keep pace with the accelerated code output, the team restructured its review process: pre-submission, Claude runs tests and assigns an independent sub-Agent for inspection; post-submission, multiple Agents concurrently identify issues, while another group performs cross-verification. Security reviews and automated testing also intercept problematic code.

Human approvals are tiered by risk level: low-risk documentation changes from trusted developers can be merged directly after passing automated checks, standard product modifications require one approval, and high-risk code demands two approvals; AI-based review cannot replace human oversight, and developers remain ultimately responsible for the final code. The team also observed that a single change underwent over 50 rounds of review and refinement, with approximately 59% of issues present from the first review. To reduce review cycles, the team now requires authors to reconsider their approach after three rounds and limits repetitive feedback.