Anthropic has been letting Claude Code handle the unglamorous side of software development — daily maintenance across iOS, Android, desktop, web, CLI, and the Agent SDK. The results are more interesting than you’d expect from an experiment Boris Cherny, the engineer behind Claude Code, describes as “surprisingly positive.”
Over a few weeks, Claude generated 388 pull requests through a dedicated Slack channel called proj-claude-maintains-apps. Human reviewers merged 180 of them — a 46 percent acceptance rate. Cherny shared the setup on X, and the prompt engineering is refreshingly boring: plain-language instructions to run crash fuzzing on iOS, Android, and desktop, use real apps without mocks, trigger crashes, and open PRs with fixes.
The routines themselves read like a maintenance team’s job descriptions. A Crash Fuzzer taps around simulators randomly to surface crashes, then traces root causes and patches them. A Dup Unifier hunts for near-duplicate abstractions and proposes merges. A Dead-Code Remover strips statically unreachable code, and when it’s unsure, it adds logging first and checks back the next day to see if anything actually executed.
A 46 percent merge rate might sound middling, but for autonomous maintenance work it’s a meaningful signal. Most PRs landed correctly on the first attempt. When they didn’t, the team adjusted the routine and Claude improved the next day — though Cherny notes that tuning sometimes takes several days. This mirrors the broader industry shift toward AI agents handling mechanical engineering tasks, but applied to a company’s own production codebase rather than a benchmark. Cherny calls it “early signs of life,” and he’s right to hedge. The real test isn’t whether Claude can open PRs — it’s whether the merged changes hold up over months of accumulated maintenance. Anthropic is now looking at ways to speed up the merge pipeline for mechanical changes, which suggests they see enough value here to keep investing.
