Claude Code now runs daily maintenance on Anthropic's software with a 46 percent merge rate

Anthropic is testing whether Claude Code can autonomously handle daily maintenance of company software. According to Boris Cherny, the creator of Claude Code, Claude has been running daily maintenance on Anthropic's internal applications over the past few weeks. The core data from this experiment: Claude created 388 pull requests, 180 of which were merged after human review, a merge rate of about 46%.

Cherny described the results as 'surprisingly positive' and considers them 'early signs of life' for the possibility of autonomous AI application maintenance. However, more than half of the auto-generated pull requests were not merged, which highlights both the potential and current limitations of this approach.

Claude runs daily routines through a dedicated Slack channel named 'proj-claude-maintains-apps', covering all of Anthropic's platforms: iOS, Android, desktop, web, CLI, and Agent SDK. Cherny shared some prompts in Slack, showing that there is no complex prompt engineering; instead, natural language instructions are used to direct Claude to start daily routines, such as crash fuzzing on iOS, Android, and desktop, using real apps rather than simulators, triggering crashes and creating fix pull requests.

Claude runs 12 specialized maintenance routines covering various aspects of code maintenance: the Crash Fuzzer opens apps in the simulator and randomly clicks to trigger crashes, analyzes root causes, and generates fixes; the Dup Unifier scans the codebase for similar but slightly different abstractions and proposes merges; the Dead-Code Remover removes statically unreachable code, and for suspicious code, first adds logging to confirm whether it is truly unused the next day. Other routines include Logic Simplifier, Logic Error Fixer, Useless Test Cleaner, Shipped Feature Inliner, Flaky Test Fixer, Abstraction Improver, and Abstraction Cop.

Pull requests created by Claude undergo a combination of automated Claude Code review and human review. Cherny stated that Claude usually creates pull requests correctly on the first try. When errors occur, the team adjusts the routines so that the AI can do better the next day, and this adjustment sometimes takes several days.

Anthropic is looking for ways to accelerate the merge process for such mechanical changes. This experiment reveals the potential of autonomous maintenance, but also shows that current human intervention and iterative tuning are needed to ensure quality.