Claude Code merges 46% of maintenance pull requests in Anthropic test

Anthropic's experiment with Claude Code represents a pragmatic exploration of how AI can absorb repetitive software maintenance tasks. Rather than a theoretical test, the company deployed Claude against its own codebase, running 12 distinct maintenance routines—from crash fuzzing to duplicate-code merging to flaky-test fixing—each with a specific role in the broader maintenance pipeline. The setup is deliberately unglamorous: plain-language Slack prompts, no complex prompt engineering, and straightforward human review gates. This approach avoids over-engineering the system while keeping humans in control of what ships.

The 46 percent merge rate after review is telling. It indicates that Claude can produce valid, useful pull requests at scale—enough to meaningfully reduce developer toil—but also that nearly half the proposals need refinement or rejection. The team's iterative approach (tweaking routines when they fail) suggests a path toward improvement over time. Cherny's framing as "early signs of life" rather than "solved" is honest and important: the experiment shows potential but does not yet prove that fully autonomous maintenance is ready for production. The fact that Anthropic is already looking at ways to accelerate the merge pipeline suggests confidence in the direction.