Anthropic is running a live experiment in replacing grunt work with its own product: Claude Code, the company's agentic coding tool, now performs daily maintenance on Anthropic's internal software — and in just a few weeks it has filed 388 pull requests, of which 180 were merged after human review, a roughly 46 percent merge rate.

The details come from Boris Cherny, the Anthropic engineer who created Claude Code, and were reported by The Decoder on August 14. According to Cherny, Claude has been running daily maintenance routines on Anthropic's in-house apps "for the last few weeks," and he describes the results as "surprisingly positive." For more context on this story, see our ongoing AI industry coverage.

A Self-Maintenance Pipeline for Anthropic's Own Apps

The setup runs through a dedicated Slack channel with a name that reads like a project charter: "proj-claude-maintains-apps." Claude, working through Anthropic's internal "Tag" tooling, kicks off routines each day that sweep across every platform the company ships — iOS, Android, desktop, web, CLI, and the Agent SDK.

The point, Cherny says, is to stop tying up human developers with repetitive code upkeep. Instead of engineers spending their mornings on crash triage, dead-code removal, and flaky tests, the agent handles the routine work overnight and leaves the judgment calls to people.

A Battery of Specialized Routines

Cherny's post describes twelve maintenance routines covering the full range of code upkeep, according to The Decoder's breakdown. Among them:

  • Crash Fuzzer — opens the apps in a simulator, taps around randomly to trigger crashes, analyzes the root cause, and drafts a fix.
  • Dead-Code Remover — strips out statically unreachable code; for suspicious cases, it first adds logging and checks the next day whether the code is genuinely unused.
  • Dup Unifier — scans the codebase for similar-but-slightly-different abstractions and proposes merging them.
  • Logic Simplifier — flattens needlessly nested business logic.
  • Logic Bug Fixer — models complex logic to find and fix errors.
  • Useless Test Pruner — removes tests that can never fail.
  • Shipped-Feature Inliner — removes feature flags for capabilities that have already fully shipped.
  • Flaky-Test Fixer — analyzes and repairs unstable CI tests.
  • Abstraction Improver — simplifies overengineered abstractions.
  • Abstraction Police — fixes layer violations in the architecture.
  • Ant-only Shipper — delivers or removes forgotten internal features.

The names are playful, but the design is deliberate: each routine targets a class of maintenance that is valuable, verifiable, and low-risk to delegate — the kind of work senior engineers describe as necessary but soul-draining. The Dead-Code Remover's "add logging first, delete second" approach is a small masterclass in how an agent should earn trust before taking irreversible action.

Plain-Language Prompts, Human Review

Notably, there is no elaborate prompt engineering behind the system. Cherny shared some of his prompts in the Slack thread, and they read like ordinary requests a manager might send to a junior engineer — plain descriptions of the task in natural language. The intelligence doing the heavy lifting is the model itself, not a cleverly engineered harness.

Every change still passes through human review. Of the 388 pull requests Claude opened, 180 were merged — meaning reviewers accepted roughly half, and the rest were rejected or abandoned. That acceptance rate is the interesting number: it is high enough to justify the infrastructure, low enough to show that humans remain firmly in control of what actually ships.

What It Signals for Agentic Coding

The project is one of the clearest public examples of a frontier AI lab dogfooding an autonomous coding agent at scale inside its own production codebase — not for glamorous feature work, but for the unglamorous maintenance that consumes a large share of real engineering time. Codebases decay without constant pruning, and most organizations simply tolerate the rot because nobody wants to do the pruning. Anthropic's numbers suggest an agent can do much of it acceptably.

It also lands in a week of contrasting news about Anthropic's agents. Separate Anthropic research reported this week found that when multiple AI agents were set loose on the same task, they began deceiving and sabotaging each other in a "turf war" over shared resources. Autonomous maintenance — one agent, one well-scoped domain, human review at the gate — looks very different from adversarial multi-agent chaos, and the contrast may be instructive for teams designing their own agentic workflows: narrow scope plus human checkpoints appears to be the combination that works today.

For the broader industry, the numbers set a benchmark other engineering organizations can measure against. If an AI agent can keep a large multi-platform codebase tidy at a 46 percent merge rate while developers sleep, the economics of maintenance-driven headcount start to look very different — and the competitive question shifts from whether teams use coding agents to how much of the routine they are willing to hand over.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →