Anthropic’s internal AI tool, Claude Code, has successfully completed 180 code merges out of 388 pull requests generated during a recent maintenance trial.
Boris Cherny, the engineer behind Claude Code, confirmed that the system has been managing daily upkeep for Anthropic’s own applications for several weeks. The results came in as “surprisingly positive” according to his post.
The setup runs via Tag through a dedicated Slack channel named “proj-claude-maintains-apps”. It targets every platform in the company’s stack, including iOS, Android, desktop, web, CLI, and the Agent SDK.
Twelve routines cover the full range of code upkeep
Cherny outlined a suite of automated tasks designed to handle repetitive code care. A “Crash Fuzzer” launches apps in a simulator, taps randomly to induce failures, identifies the root cause, and drafts a fix. A “Dup Unifier” scans the codebase for similar but distinct abstractions and suggests combining them. The “Dead-Code Remover” strips out unreachable code but inserts logging first to verify if the code is truly unused over the next day.
Other routines include an “Ant-only Shipper” that delivers or removes forgotten internal features, a “Logic Simplifier” that untangles nested business rules, and a “Logic Bug Fixer” that models complex logic to locate errors. There is also a “Useless Test Pruner” that deletes tests incapable of failing and a “Shipped-Feature Inliner” that removes flags for fully released features. The “Flaky-Test Fixer” addresses unstable CI tests, while an “Abstraction Improver” simplifies overengineered structures. Finally, an “Abstraction Police” corrects violations of the architectural layering.
Cherny shared some of these prompts in Slack, noting there is no elaborate prompt engineering involved. He instructs the AI in plain language to start daily routines for crash fuzzing on iOS, Android, and desktop. The tool uses real apps without mocks, triggers crashes, and creates pull requests containing the fixes.
388 pull requests created, just under half merged
The system generated 388 pull requests across Anthropic’s repositories over the first few weeks. After automated review by Claude Code and human oversight, 180 were merged, representing a rate of about 46 percent.
Cherny noted that Claude Code usually gets the pull requests right on the first attempt. When it does not, the team tweaks the routine so the AI performs better the next day. That tuning sometimes takes a few days. Anthropic is now exploring ways to speed up the merge process for these mechanical changes.
More than half of the auto-generated pull requests did not make it through, which highlights both the limits and the potential of the approach. Cherny describes the experiment as “early signs of life” that autonomous AI-powered app maintenance could work.
What it means
For developers, this trial shifts the burden of repetitive maintenance away from human teams. The AI handles the heavy lifting of finding crashes, cleaning up dead code, and fixing flaky tests. While the 46 percent merge rate shows the tool is not perfect yet, it proves the software can operate independently on a daily basis. The team is now focused on refining the process to get more of these changes accepted without manual intervention.
