Anthropic is using its own coding agent to see how far AI can go in routine software upkeep. Claude Code has been running daily maintenance on Anthropic's in-house apps for "the last few weeks," according to Boris Cherny, the Anthropic engineer who created Claude Code.
The experiment is not replacing review. Claude Code creates pull requests, but changes still go through automated Claude Code review and human review before they are merged.
What Anthropic Is Testing
The core question is practical: can Claude Code take on repetitive maintenance work across real software products without constant developer involvement?
According to Cherny, Claude is already handling daily routines on its own. The work runs through a dedicated Slack channel called "proj-claude-maintains-apps" and covers Anthropic's platforms: iOS, Android, desktop, web, CLI, and the Agent SDK.
That scope matters because daily maintenance is rarely limited to one surface. Bugs, unused code, duplicated abstractions, and platform-specific crashes can appear across many parts of a software stack. Anthropic is testing whether an AI coding agent can keep moving through that kind of work as a repeatable process.
The Maintenance Jobs Claude Code Runs
Cherny describes twelve routines that cover a range of code upkeep tasks. The examples show a workflow aimed at finding small but persistent problems, then turning them into pull requests developers can review.
One routine is called "Crash Fuzzer." It opens apps in a simulator, taps around randomly to trigger crashes, analyzes the root cause, and creates a fix. The important point is that the process uses real apps without mocks, according to the prompt Cherny shared.
Another routine is "Dup Unifier." It searches for abstractions that are similar but not quite the same, then proposes merging them. This is the sort of cleanup that can reduce long-term code complexity, but it still needs careful review because similar code is not always interchangeable.
A third routine is "Dead-Code Remover." It removes statically unreachable code. When code looks suspicious rather than clearly unused, Claude first adds logging, then checks the next day whether the code is actually unused.
These examples point to a pattern: Claude Code is being used where maintenance work can be broken into repeatable checks, evidence gathering, fixes, and pull requests.
The Numbers So Far
In the first few weeks, Claude Code created 388 pull requests across Anthropic's repositories, Cherny says. Of those, 180 were merged after automated Claude Code review and human review. That works out to a rate of about 46 percent.
That number cuts both ways. A 46 percent merge rate suggests Claude Code is producing enough useful work to matter. At the same time, more than half of the generated pull requests did not make it through.
For software teams, that distinction is central. An AI maintenance system is valuable only if it reduces the total burden on engineers. A steady stream of pull requests can help if many are correct, but it can also create review load if the work is noisy or poorly scoped.
Cherny says Claude usually gets pull requests right on the first try. When it does not, the team adjusts the routine so the AI performs better the next day. That tuning can sometimes take a few days.
Why Review Still Matters
The experiment keeps human review in the loop, which is a key part of the workflow described. Claude Code may find issues and propose fixes, but merged changes still require review.
That structure makes the system less like a fully unsupervised developer and more like an automated maintenance contributor. It can search for problems, make changes, and submit work, while humans remain responsible for accepting the result.
The use of automated Claude Code review before human review also shows how Anthropic is layering checks. The source does not describe those review steps in detail, but it does state that both automated and human review are part of the merge path.
Anthropic is now looking at ways to speed up the merge process for these kinds of mechanical changes. That suggests the bottleneck may not only be code generation, but also the process of deciding which generated changes are safe and useful enough to merge.
What This Signals For AI Coding Agents
Cherny calls the results "surprisingly positive" and describes the experiment as "early signs of life" for autonomous AI-powered app maintenance.
The phrase is cautious for good reason. The current results show potential, but they also show limits. Claude Code created hundreds of pull requests, yet only 180 were merged. The system is productive, but it is not automatically correct.
Still, the experiment is a concrete example of how AI coding agents may move beyond one-off prompts. Instead of waiting for a developer to ask for a single fix, Claude Code is being assigned recurring maintenance routines across active software platforms.
If that model keeps improving, the role of AI in software teams could become more operational: checking for crashes, identifying duplication, removing dead code, and preparing pull requests for review. Based on Anthropic's test, the near-term value is not magic autonomy. It is steady, reviewable maintenance work that developers can accept, reject, or tune for the next run.