Anthropic is using Claude Code to perform daily maintenance on its own software, according to The Decoder, which cites Boris Cherny, the Anthropic engineer who created Claude Code. The experiment has been running for “the last few weeks,” with Claude generating pull requests for repetitive upkeep across Anthropic repositories. The reported result so far: 388 pull requests created by Claude Code in the first few weeks, with 180 merged after a combination of automated Claude Code review and human review. That works out to about a 46% merge rate — high enough to show useful output, but also a reminder that more than half of the AI-generated changes did not make it through. The workflow described by The Decoder is operational rather than demo-oriented. Claude runs maintenance routines through a dedicated Slack channel called “proj-claude-maintains-apps,” using a tool called Tag, across Anthropic’s iOS, Android, desktop, web, command-line interface, and Agent SDK platforms. The work is aimed at routine software hygiene rather than new product features. The routines include a “Crash Fuzzer,” which opens apps in a simulator, taps around to trigger crashes, analyzes the cause, and creates a proposed fix. Another routine, “Dup Unifier,” looks for similar but slightly different abstractions in the codebase and proposes consolidating them. A “Dead-Code Remover” strips out statically unreachable code; when code looks suspicious rather than clearly unused, it first adds logging so the system can check the next day whether that code is actually unused. According to The Decoder’s account of Cherny’s post, the prompts are plain-language instructions rather than elaborate prompt-engineering artifacts. Cherny tells Claude to start daily crash-fuzzing routines on iOS, Android, and desktop, use real apps rather than mocks, trigger crashes, and create pull requests with fixes. The review loop is still central. The Decoder reports that Claude often gets pull requests right on the first try, according to Cherny, but when it does not, Anthropic adjusts the routine so the system performs better the next day. That tuning can take several days. The most important constraint is the merge gate. A 46% merge rate indicates that Claude Code is producing enough acceptable work to be useful inside Anthropic’s own engineering process, but not enough to remove human review from the loop. The Decoder says Anthropic is now looking at ways to speed up the merge process for these mechanical changes. Who benefits: Anthropic’s internal engineering teams benefit if Claude Code can reduce repetitive maintenance work across apps. Claude Code also gains credibility if Anthropic can show sustained internal use with merged changes. Who's exposed: Reviewers remain exposed to quality-control burden because more than half of the generated pull requests were not merged. Teams adopting a similar process would need to account for tuning time and review capacity, not just agent output.