Anthropic says Claude now writes about 90 percent of the Claude Code codebase, forcing the team to redesign review for a repository receiving more than 30,000 merged changes a day. Its target is for most pull requests to merge with no more than one human involved and without increasing incidents or reversions.
The team says nobody maintains a complete mental model of the product. When engineers need to understand one area, they ask Claude to generate a current explainer for that slice of code. One documented pull request went through more than 50 review rounds overnight, while automated reviewers also spent substantial effort identifying problems in code that existed before the proposed change.
That approach has prompted a useful objection from outside practitioners: if the system fails in a way the model cannot explain, there may be no person with enough causal understanding to debug it methodically. Large legacy codebases already have fragmented ownership, but deliberately depending on generated explanations changes the trust model.
Anthropic’s experience demonstrates throughput, not yet long-term maintainability. The important measures will be incident rates, repair time, reversions and whether independent reviewers can reproduce an explanation. High merge volume is valuable only if the organization can still identify why the software behaves as it does when automation is wrong.