Claude opened 388 pull requests on Anthropic's own apps. 180 got merged.

Anthropic has spent the last few weeks letting Claude Code perform daily maintenance on its own in-house apps, and there is now a number attached to the result: 388 pull requests opened, 180 merged ^[1][2][3][5]. That works out to a merge rate of about 46 percent ^[4], which is the first concrete figure operators have for what happens when routine upkeep is handed to an agent.

The number cuts both ways, and the arithmetic is the interesting part. If 180 of 388 landed, 208 did not ^[16], roughly 1.2 discarded pull requests for every one that shipped ^[18]. Every one of those still consumed a review slot: the merge decision came after a combination of automated Claude Code review and human review, according to Boris Cherny, the Anthropic engineer who created Claude Code ^[4][5]. The cost of this arrangement is not inference, it is reviewer attention, and at a 46 percent hit rate the agent generates slightly more review work that ends in nothing than review work that ends in a merge. Anthropic appears to know it: Cherny says the team is looking at ways to speed up the merge process for mechanical changes ^[14].

The setup is less exotic than the output suggests. Claude runs through a dedicated Slack channel called "proj-claude-maintains-apps," via a tool Cherny refers to as Tag, across iOS, Android, desktop, web, CLI, and the Agent SDK ^[7]. Twelve routines cover the upkeep surface ^[8]. A Crash Fuzzer opens apps in a simulator, taps around at random to provoke a crash, analyses the root cause and writes a fix ^[9]. A Dup Unifier looks for near-identical abstractions and proposes merging them ^[10]. A Dead-Code Remover deletes statically unreachable code, and for code it is only suspicious about, it first adds logging and checks the next day whether the path is actually cold ^[11]. That last one is the detail worth copying: the routine is designed to defer its own risky decisions by a day rather than guess.

Cherny also says there is no elaborate prompt engineering behind any of it, and shared prompts written in plain language instructing Claude to start daily crash-fuzzing routines on iOS, Android and desktop, use real apps rather than mocks, trigger crashes, and open pull requests with fixes ^[12]. The tuning loop lives at the routine level, not the prompt level: Claude usually gets a pull request right first time, and when it does not, the team adjusts the routine so the next day goes better, which sometimes takes a few days ^[13]. Cherny calls the results "surprisingly positive" and the experiment "early signs of life" for autonomous app maintenance ^[6][15].

Treat all of this as single-sourced vendor self-reporting on a vendor product. Worth watching: the denominator behind "a few weeks," since the source gives no exact duration and therefore no daily pull request volume ^[5]; whether the 46 percent holds as routines move past mechanical deletions and deduplication; the revert rate on the 180 that merged, which nobody has published; and what the faster merge path for mechanical changes actually removes from review ^[14].