Anthropic has turned its internal software maintenance over to Claude Code, treating its own repositories as a proving ground for autonomous development. Over several weeks of operational testing, the coding agent authored 388 pull requests, with human engineers approving and merging 180 of them—landing a 46% conversion rate. According to Boris Cherny, the Anthropic engineer behind Claude Code, the initiative aims to unburden developer teams from grinding maintenance cycles across production stacks.
Operating through an internal Slack channel dubbed "proj-claude-maintains-apps" via Tag, the workflow orchestrates twelve specialized maintenance jobs spanning Anthropic’s iOS, Android, web, desktop, CLI, and Agent SDK platforms. The setup runs on plain-English instructions rather than bespoke prompt engineering. Key pipelines include a "Crash Fuzzer" that generates random input sequences to reproduce edge-case crashes and propose patches, a "Dup Unifier" targeting redundant abstractions, and a "Dead-Code Remover" that instruments telemetry checks before wiping unused code.
The 46% acceptance figure sketches the realistic ceiling of enterprise agentic workflows today. Claude Code successfully eliminates mechanical drafting, but more than half of its proposed diffs fail human review, shifting senior engineering hours from writing boilerplate to filtering synthetic pull requests. Autonomous maintenance works in production, but the economic math holds only as long as engineering leads are willing to review the 54% that misses the mark.