Claude Code now runs daily maintenance on Anthropic's software with a 46 percent merge rate
| Source: THE DECODER
Tags: Claude Code, Anthropic, autonomous coding, agentic AI, Boris Cherny, software maintenance
Claude Code has been autonomously running daily maintenance on all of Anthropic's apps for several weeks, generating 388 pull requests with a 46% merge rate after human review — concrete production-scale evidence that agentic coding can clear a quality bar in real codebases.
Details
Boris Cherny, the Anthropic engineer who built Claude Code, reports that the tool has been running daily automated maintenance on all of Anthropic's in-house apps for several weeks. Claude operates through a dedicated Slack channel called proj-claude-maintains-apps, executing 12 distinct maintenance routines across iOS, Android, desktop, web, CLI, and the Agent SDK. The routines span the full range of code upkeep: a Crash Fuzzer opens real apps in simulators and taps randomly to trigger crashes then proposes root-cause fixes; a Dead-Code Remover strips unreachable code but adds logging first for suspicious cases to verify the next day; a Flaky-Test Fixer analyzes unstable CI tests; a Dup Unifier proposes merging similar abstractions; a Shipped-Feature Inliner removes feature flags after full rollout. Critically, the prompts driving Claude are plain-language Slack messages — no elaborate prompt engineering. Of 388 pull requests created, 180 (46%) were merged after automated Claude Code review plus human review. Cherny calls results 'surprisingly positive' and frames it as 'early signs of life' for agentic maintenance — deliberately modest language that makes the metric more credible, not less. The 46% rate sits well below rubber-stamping territory while demonstrating real autonomous utility at production scale.