Anthropic is testing whether Claude Code can handle routine maintenance across the company’s own software, The Decoder reports.
The internal experiment covers tasks such as crash fuzzing and dead-code removal. Over a few weeks, Claude Code created 388 pull requests, and 46 percent were merged after human review. Boris Cherny, the inventor of Claude Code, described the results as early signs that this kind of workflow may be possible.
The numbers are useful because they place agentic coding in a realistic setting. A merged pull request is a stronger signal than a demo, but the human review requirement shows that the agent is not operating as an independent maintainer. People still decide which changes are safe, useful, and worth shipping.
For engineering teams, the likely near-term value is in the backlog of small improvements that are easy to postpone. If agents can prepare plausible fixes and cleanup patches, maintainers may spend more time reviewing and prioritizing, and less time doing repetitive mechanical work from scratch.