Anthropic engineer Boris Cherny let an internal deployment of Claude Code operate largely unsupervised on the company’s own applications for several weeks, generating pull requests without an engineer requesting each one. The AI opened 388 pull requests spanning Anthropic’s iOS, Android, desktop, web, CLI, and Agent SDK codebases, and reviewers approved 180 of them, a 46 percent merge rate. That number cuts two ways: it took a machine, not a person, to produce 388 candidate fixes with no one prompting each individual change, and it also means reviewers turned back the majority of what it proposed.
Cherny, who created Claude Code, described the setup in a post he shared internally at Anthropic. Claude connects through Tag, Anthropic’s tool for running Claude inside Slack, and works from a Slack channel Cherny named “proj-claude-maintains-apps.” He told the model in plain language, no structured prompt engineering, to open the apps, hunt for crashes, simplify tangled logic, and clear out code nobody uses anymore.
The routines fall into roughly a dozen categories, each aimed at a specific kind of upkeep:
- Crash Fuzzer, which loads an app in a simulator, taps around at random to trigger a crash, and drafts a fix once it finds one
- Dup Unifier, which flags near-duplicate code abstractions and suggests merging them
- Dead Code Removal, which strips unreachable code and logs suspicious cases before deleting them
- Flaky-Test Fixer, which diagnoses and repairs unstable tests in continuous integration
Claude tends to get a pull request right on its first attempt, Cherny says. When it does not, the team adjusts the routine’s instructions rather than hand-editing the code, a tuning process that can stretch over several days. Anthropic is now exploring ways to move these mechanical changes through review faster.
The 46 percent figure is the headline number, but nothing in the report says how many reviewer hours the 208 rejected pull requests cost. If waving off a bad PR takes thirty seconds, an agent that proposes 388 changes and gets under half of them right is a clear net gain for the team. If each rejection demands the same scrutiny as reviewing a colleague’s code, the arithmetic looks far less favorable. Cherny frames the results as “early signs of life,” his own phrase for evidence that autonomous maintenance might eventually work. That is a hypothesis, not a completed measurement.
The experiment also says less about agentic coding tools in general than it does about Anthropic running its own model against its own repositories, reviewed by engineers who already know the model’s habits, on maintenance work chosen precisely because it is low stakes. That is closer to the most forgiving test an AI coding agent could face than to a production deployment at an unfamiliar company. Extending this merge rate to teams running Claude Code, or any coding agent, against systems the model has never seen requires evidence this report does not provide.
The finding lands inside a debate AI Insiders has been tracking this week: a three-person team merging hundreds of pull requests weekly credited its output to adversarial review, not autonomous agent loops, and a separate argument holds that verification capacity, not model capability, is now the binding constraint on agent adoption. Anthropic’s 46 percent merge rate is a fresh data point in that argument. It is not a resolution of it.
Engineering leaders piloting agentic maintenance tools should track the reviewer hours spent on rejected pull requests, not just the headline merge rate, before deciding whether the trade actually saves anyone time.
The Decoder, a German technology publication, reported this story on 14 August 2026.
