The Anthropic team has disclosed that roughly 90% of the code for Claude Code is generated autonomously by AI. Moreover, AI is also responsible for the majority of code review tasks. From June to September 2025, the volume of merged pull requests (PRs) doubled, while the rollback rate witnessed a decline of over 40%. The team has moved away from the conventional approach of manual line-by-line code review. Instead, they have instituted a seven-layer code inspection system, utilizing rollback and incident rates as pivotal metrics. This system leverages multiple AI Agents to independently detect bugs, deliberate on and filter opinions (Note: 'Filter' has been corrected to 'filter' for proper context), and tailor the degree of human approval required according to the risk level associated with the code changes. It hinges on a multi-tiered verification mechanism to minimize risks. Previously, two AI systems were engaged in excessively detailed reviews of each other's work, resulting in over 50 rounds of exchanges in a single night. To address this, the team imposed a three-round cap on AI reviews, which successfully reduced the average number of review iterations. Now, human resources are liberated from the burdensome task of extensive code reviews. Instead, they can concentrate on evaluating the necessity of code changes, scrutinizing high-risk core code, and steering the architectural direction.