Anthropic Tests Claude Code for Daily Software Maintenance

Anthropic has begun using its own Claude Code tool to handle daily software maintenance tasks across its entire product ecosystem, including iOS, Android, desktop, web, CLI, and the Agent SDK. Created by Anthropic engineer Boris Cherny, the system operates through a dedicated Slack channel called "proj-claude-maintains-apps." Using plain-language prompts rather than complex prompt engineering, the AI autonomously executes 12 specialized maintenance routines to keep the company's codebases clean and functional.

During the first few weeks of this internal trial, Claude Code generated 388 pull requests. Following a combination of automated reviews and human oversight, 180 of these requests were successfully merged, representing a 46 percent merge rate. The AI's routines include a "Crash Fuzzer" that runs apps in simulators to trigger and fix crashes, a "Dup Unifier" that merges redundant code abstractions, and a "Dead-Code Remover" that deletes unreachable code or logs suspicious code to verify if it is unused. It also removes obsolete feature flags and resolves architectural layer violations.

For software practitioners, this experiment demonstrates the viability of delegating tedious, repetitive upkeep to autonomous agents. When Claude Code fails to get a pull request right on the first attempt, developers can fine-tune the routine over a few days to improve future performance. By automating these mechanical chores, engineering teams can shift their focus from routine debugging and code cleanup to building new features, signaling a shift toward highly autonomous codebase maintenance.