Claude Code Hits 46% Merge Rate Running Daily Maintenance at Anthropic
Agentic AIA 46% merge rate isn’t a demo — it’s a production signal. Anthropic is running Claude Code as a daily maintenance system across its own iOS, Android, desktop, web, and CLI products, and nearly half of every pull request it generates is surviving human review.
What Happened
Engineer Boris Cherny built a system that delivers plain-language Slack instructions to Claude Code each day, triggering one of twelve automated maintenance routines against Anthropic’s own production codebases. Over several weeks of operation, Claude generated 388 pull requests. Human and automated review accepted 180 — a 46% merge rate that Cherny described as “surprisingly positive.”
The twelve routines are concrete and unglamorous: crash fuzzing, dead code removal, duplicate consolidation, logic simplification, feature flag cleanup, flaky test repair, and architecture violation detection. When a PR fails review, Cherny’s team adjusts the routine specification — not the model — for the next iteration. The system is designed to be tunable, treating each failure as input rather than an exit condition.
Anthropic is also reportedly exploring ways to accelerate the review and merge process for mechanical changes, suggesting the active bottleneck is now review throughput, not generation quality.
Why It Matters
The 46% figure matters because of the denominator. Claude isn’t running against a cherry-picked corpus of toy examples — it’s operating on Anthropic’s live production code across multiple platforms. A nearly one-in-two acceptance rate on autonomous, unsupervised maintenance work is the closest thing to a real production benchmark the industry has produced for coding agents.
What the 12-routine taxonomy reveals is equally significant: every task type is systematically non-urgent. No one is blocked by a stale feature flag. A flaky test isn’t a P1 incident. Dead code isn’t stopping a release. This is precisely the category of work that human engineers perpetually defer — the technical debt accumulation that compounds until it demands a cleanup sprint no one wants to own. Claude Code doesn’t deprioritize these tasks. It shows up every day.
The 12-routine framework also has an important architectural implication: tasks are scoped and typed before execution. Claude isn’t given a vague prompt to “improve things.” It receives a parameterized routine with a defined target category (e.g., “remove dead code from the iOS app”) and executes against that constraint. This is the same pattern that distinguishes reliable agentic workflows from unpredictable open-ended runs — as explored in the economics of running Claude at production scale.
What This Means for Builders
The system Cherny built has three components that are fully replicable outside Anthropic: a typed routine library, a delivery mechanism, and a human-in-the-loop review step. None of this requires internal tooling. The same pattern — parameterized routines triggered on a schedule, gated by review before merge — is reproducible with any coding agent and any CI system that supports PR-based workflows.
The operational implication is that the right framing for autonomous AI maintenance is not “let the agent improve the codebase” but “define the category of work, then let the agent execute within it.” Teams that invest in the routine library — the classification of what kinds of maintenance work exist and what a correct output looks like — get a compounding return. Teams that skip this step get noise.
The So What
Claude Code running daily maintenance at Anthropic with a 46% merge rate is the benchmark the industry needed but didn’t have for autonomous coding agents. It shifts the conversation from “can AI write code?” — settled — to the more useful question: “what specific maintenance categories can AI cover reliably enough that review overhead is worth it?”
At Anthropic’s scale, the answer is dead code, duplicate logic, flaky tests, feature flags, and architecture drift. The experiment also surfaces where human attention is still required: not in generation, but in routine design and merge throughput. That’s a different resourcing problem than the one most teams are currently solving for.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →