Claude Code runs daily maintenance for Anthropic with 46% PR merge rate
Anthropic has begun letting Claude Code run daily maintenance on its own software, from crash fuzzing to dead-code removal. In a few weeks, the agent produced 388 pull requests, and about 46% were merged after human review—a signal that autonomous coding agents are moving from demos to dependable operations. The experiment, led by Claude Code's creator Boris Cherny, suggests routine upkeep can be delegated while engineers focus on higher-level work. For builders, this offers a concrete benchmark for AI reliability in continuous integration, though more than half of the proposals were rejected, underscoring that human oversight remains essential. The result points to a near-term future where AI agents handle recurring chores, with teams refining prompts and review workflows to raise acceptance rates.
Claude Code runs daily maintenance for Anthropic with 46% PR merge rate
Anthropic is testing Claude Code for daily maintenance of its own apps. In a few weeks, the AI created 388 pull requests, and 46 percent were merged after human review.
Key takeaway
A 46% merge rate for AI-generated maintenance PRs marks a critical inflection point where coding agents shift from novelty to practical utility in everyday software operations.
What happened
Anthropic is testing whether Claude Code can handle daily maintenance of its own apps, from crash fuzzing to dead-code removal. Over a few weeks, the AI generated 388 pull requests, with 46% successfully merged after human review, according to Claude Code inventor Boris Cherny.
The agent runs daily routines through a dedicated Slack channel called 'proj-claude-maintains-apps' across Anthropic's platforms, using tools like a Crash Fuzzer that taps apps to trigger failures, a Dup Unifier to merge similar abstractions, and a Dead-Code Remover that first adds logging for suspicious code. Cherny described the results as 'surprisingly positive' and sees 'early signs of life' for autonomous maintenance.
Evidence
Anthropic is testing Claude Code for daily maintenance of its own apps, including crash fuzzing and dead-code removal.
The Decoder · attributed
Anthropic is testing whether Claude Code can handle daily maintenance of the company's own apps, from crash fuzzing to dead-code removal.
Claude Code created 388 pull requests and 46% were merged after human review.
The Decoder · attributed
Claude created 388 pull requests across Anthropic's repositories in the first few weeks, Cherny says. After a combination of automated Claude Code review and human review, 180 were merged, a rate of about 46 percent.
Claude Code runs daily maintenance routines via a Slack channel across Anthropic's platforms, using tools like Crash Fuzzer, Dup Unifier, and Dead-Code Remover.
The Decoder · attributed
Claude, via Tag, runs daily routines through a dedicated Slack channel called 'proj-claude-maintains-apps' across all of Anthropic's platforms: iOS, Android, desktop, web, CLI, and the Agent SDK. Claude runs a battery of specialized maintenance routines, according to Cherny. A 'Crash Fuzzer' opens apps in a simulator, taps around randomly to trigger crashes, analyzes the root cause, and creates a fix. A 'Dup Unifier' scans the codebase for similar but slightly different abstractions and proposes merging them. The 'Dead-Code Remover' strips out statically unreachable code and, for suspicious code, first adds logging so it can check the next day whether the code is actually unused.
Anthropic engineer Boris Cherny describes the results as 'surprisingly positive' and sees 'early signs of life' for autonomous maintenance.
The Decoder · attributed
Cherny calls the results 'surprisingly positive.' ... Cherny describes the experiment as 'early signs of life' that autonomous AI-powered app maintenance could work.
Why it matters
For engineering teams, this validates that delegating repetitive maintenance to AI agents is viable, but also exposes the need for automated review and prompt refinement to boost acceptance rates beyond half.
Limits and uncertainties
The merge rate is below half, indicating that a significant amount of human review and rejection still occurs.
The experiment is limited to mechanical maintenance tasks, not complex feature development or architectural changes.
The tuning process sometimes takes days, suggesting that prompt engineering is not yet fully automated.
Practical implications
Engineering teams can adopt AI agents for routine maintenance, but should invest in automated review and prompt tuning to improve merge rates.
Integrating agents into CI/CD pipelines requires careful setup, monitoring, and a dedicated feedback loop for refining routines.
Developers should track merge rates and iteratively adjust routines based on failures.
What to watch
Watch for further improvements in merge rates and expansion of maintenance tasks beyond mechanical changes.
Observe whether Anthropic and others publish more details on the tuning process and failure analysis.
