Editorial illustration for Claude Code now runs daily maintenance on Anthropic's software
Claude Code Now Runs Daily Maintenance Unsupervised
Anthropic has started letting its own AI coding tool loose on its own codebase, unsupervised, every day. Boris Cherny, the engineer who built Claude Code, says the model has spent the last few weeks running daily maintenance across the company's software: iOS, Android, desktop, web, CLI, and the Agent SDK. The numbers so far: 388 pull requests generated, 180 merged after human review, a 46 percent hit rate that Cherny calls "surprisingly positive."
The setup runs through a Slack channel named "proj-claude-maintains-apps," where Claude, working through a system Anthropic calls Tag, checks in on a dozen distinct maintenance routines without a developer prompting each run. The goal is to take the grunt work off engineers' plates, the crash-chasing and cleanup nobody wants to do by hand, and see if an AI can do it reliably enough to trust with production code. Cherny has been posting some of the actual prompts he uses, and they're plain, almost conversational instructions rather than anything engineered for the occasion. What those routines actually look like, and how far Anthropic is willing to let Claude run without a human catching mistakes, is where it gets interesting.
Anthropic is testing whether Claude Code can handle daily maintenance of the company's own software. In a few weeks, the AI created 388 pull requests, and 180 were merged after human review.
Why this matters
A 46 percent merge rate on 388 machine-generated pull requests is a real number, not a demo reel, and that's what makes this worth watching. Cherny didn't need clever prompting tricks; he told Claude in plain English to fuzz-test real apps and submit fixes, and it did that daily for weeks. For developers, the interesting part isn't that Claude wrote code, it's that the code passed human review often enough to matter for unglamorous, repetitive work like crash testing across iOS, Android, and desktop.
Founders running lean engineering teams should read this as a signal that this class of maintenance, the stuff nobody wants on their sprint board, is a plausible place to offload hours now, not eventually. Researchers should note the other half of that number too: 208 pull requests didn't survive review. Anthropic is still keeping a human in the loop, and the honest framing here, tested internally on its own products before anyone else's, is exactly the kind of disclosure the rest of the industry should be matching.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives Liquid AI's 3B Vision Model Shows Major Gains in Screen Reading, Object Grounding
From Archives OpenAI Slows AI Model Releases, Details Safety Shortcomings
From Archives Claude Agents Sabotaged Each Other on Shared Server, Hid Actions From Users
From Archives Anthropic's AI Agents With Incompatible Goals Started a Turf War
Research & Benchmarks
Common Questions Answered
What is Claude Code's merge rate for the daily maintenance pull requests at Anthropic?
Claude Code achieved a 46 percent merge rate on 388 pull requests generated over a few weeks of daily maintenance work. This means 180 pull requests were successfully merged after human review, which Boris Cherny, the engineer who built Claude Code, describes as surprisingly positive for an unsupervised AI system.
Which software platforms has Claude Code been maintaining for Anthropic?
Claude Code has been running daily maintenance across multiple Anthropic software platforms including iOS, Android, desktop, web, CLI, and the Agent SDK. The AI system has been handling maintenance tasks across all these different codebases simultaneously over the course of several weeks.
How does Claude Code receive its daily maintenance instructions at Anthropic?
Claude Code receives its daily maintenance instructions through a Slack channel and operates unsupervised, running automatically each day. Boris Cherny noted that he didn't need to use clever prompting tricks, but simply told Claude in plain English to fuzz-test real apps and submit fixes, which it then executed daily for weeks.
Why is Anthropic's 46 percent merge rate on Claude Code's pull requests significant?
The 46 percent merge rate represents real-world performance rather than a controlled demo, making it meaningful for practical development work. This demonstrates that Claude Code can handle unglamorous, repetitive tasks like crash testing across multiple platforms at a rate high enough to provide genuine value for developers without requiring extensive manual intervention.
Further Reading
- We tested Anthropic's redesigned Claude Code desktop app and Routines — here's what enterprises should know - VentureBeat
- Claude Code Routines: Anthropic's Answer to Unattended Dev Automation - DevOps.com
- How Anthropic engineering teams use Claude Code every day - Codingscape
- Schedule recurring tasks in Claude Code Desktop - Anthropic Docs
- Advanced setup - Claude Code Docs - Anthropic Docs
