Claude Code Begins Using AI to Review AI: 90% of Code Written by Claude, Low-Risk Changes Require No Human Approval

Beating AI Newsflash: Anthropic has revealed that approximately 90% of the code for Claude Code is now written by Claude. From June to September this year, the number of pull requests (PRs) merged by the Claude Code team more than doubled, while the proportion of rollbacks decreased by over 40%. To keep up with the speed of AI-generated code, Anthropic has restructured its code review process. Before each submission, Claude runs tests and has independent sub-agents review the modifications. After submission, multiple agents work in parallel to identify bugs, followed by cross-verification by another group of agents. Security reviews and automated testing also prevent problematic code from being merged. Human reviews are tiered based on risk levels. Low-risk changes, such as documentation updates submitted by trusted developers, can be merged directly after passing automated checks. Standard product changes require approval from one person, while high-risk code changes need approval from two people. AI reviews cannot replace human oversight, and developers remain responsible for the final code. However, AI reviews can sometimes fall into a feedback loop. Anthropic disclosed that one Claude Code modification underwent over 50 rounds of review and fixes overnight. Upon investigation, the team found that approximately 59% of the issues stemmed from code that existed during the first round of review, rather than new problems introduced by subsequent changes. The team then adjusted the rules, requiring authors to reconsider their approach after more than three rounds of review and limiting repetitive feedback to prevent agents from getting entangled with one another. [Original Link]