# Anthropic’s Claude opened 388 maintenance pull requests, with 46% merged after review > Source: > Published: 2026-08-14 15:05:01.346160+00:00 # Anthropic’s Claude opened 388 maintenance pull requests, with 46% merged after review - Claude opened 388 pull requests across Anthropic’s iOS, Android, desktop, web, CLI and Agent SDK repositories; 180 were merged after automated and human review. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) - The routines cover crash fuzzing, dead-code removal, flaky-test repair, duplicate-code cleanup, logic simplification and abstraction fixes. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) - Anthropic says it tunes routines after failures, but has not disclosed line counts, review hours, post-merge defects or why each of the 208 unmerged pull requests failed. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) - An external study found Claude Code pull requests merged at a lower rate than human-authored pull requests in its sample, but the result is not directly comparable with Anthropic’s internal maintenance experiment. [[4]](https://arxiv.org/abs/2509.14745) Anthropic is testing whether Claude can keep its software in shape with a daily stream of automated maintenance tasks. Over several weeks, routines running through Claude Tag opened 388 pull requests across the company’s repositories; 180 were merged after Claude Code Review and human review, Boris Cherny, the creator and head of Claude Code, said in a LinkedIn post published August 14, 2026. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) The experiment covers Anthropic’s iOS, Android, desktop and web apps, as well as its command-line interface and Agent SDK. Cherny described the result as “early signs of life” for autonomous software maintenance, while saying the company is still working out how to make review and merging faster. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) ## The routines target work engineers often postpone The system operates from a Slack channel called “proj-claude-maintains-apps.” One routine launches an app in a simulator, taps through it to trigger crashes, investigates the cause and proposes a fix. Another searches for similar but divergent abstractions and opens pull requests to unify them. A dead-code routine removes code that static analysis identifies as unreachable; where the result is less certain, it adds logging first and waits for evidence before deleting the code. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) Anthropic’s list also includes routines for simplifying nested business logic, fixing logic bugs, pruning tests that can never fail, removing feature flags from fully shipped features, repairing flaky tests and correcting architectural layer violations. Cherny said the prompts are generally written in plain language rather than built around elaborate prompt engineering. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) The task selection matters. Crash reproduction, unreachable-code checks and flaky-test diagnosis offer relatively clear signals for success. Abstraction changes and logic fixes require more judgment about design intent, and Anthropic did not publish a breakdown showing which categories produced the accepted or unmerged pull requests. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) ## Review remains the gate Cherny’s account says the 180 accepted pull requests passed Claude Code Review and human review. Anthropic’s Claude Code team has separately described a layered process: an automated review bot examines every pull request, while code owners manually review changes to critical areas. For less critical outer layers, Anthropic is increasingly relying on automated review after a multimonth effort to establish confidence. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/)[[2]](https://simonwillison.net/2026/Jul/21/cat-and-thariq/) Cherny said Claude generally gets these maintenance pull requests right on the first attempt. When it does not, Anthropic tunes the routine so it performs better the following day; some adjustments take several days. That describes an iterative workflow in which the company improves task instructions and validation after failures, rather than treating each pull request as an isolated model output. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) The company has not disclosed the code volume behind the 388 pull requests. There is no reported figure for lines added or removed, files changed, token use, reviewer time or post-merge defects. The 46% figure therefore measures pull-request acceptance, not productivity or software quality. The remaining 208 pull requests were not described individually, so the public account does not establish how many were rejected for technical reasons, duplicates, low priority, inactivity or other causes. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) ## The result is promising, but not a benchmark Available research shows why Anthropic’s 46% figure should be treated as an early operating metric rather than a scorecard for AI versus human engineers. A study of 567 Claude Code pull requests across 157 open-source projects found that 83.8% were eventually accepted and merged, compared with 91.0% for human-authored pull requests in the study’s comparison sample. The authors also found that 54.9% of merged agent pull requests were integrated without further modification. [[4]](https://arxiv.org/abs/2509.14745) That study involved public repositories and a broader mix of refactoring, documentation, testing and feature work. Anthropic’s experiment focuses on internally selected maintenance routines, and its denominator is only a few weeks of company-specific pull requests. The two rates therefore should not be read as a direct comparison between Anthropic’s system and human engineers. [[1]](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/)[[4]](https://arxiv.org/abs/2509.14745) A separate study of 3,225 agent-generated fix pull requests from the AIDev dataset found that 46.41% were rejected. The dataset combined contributions from Copilot, Devin, Cursor and Claude across open-source projects. Researchers identified rejection reasons including incorrect implementations, failed tests or continuous integration, inability to complete the task, inactivity, superseded changes and low-priority fixes. Its roughly 53.6% acceptance rate is closer to Anthropic’s 46% merge rate, but the datasets, task definitions and review practices still differ. [[5]](https://arxiv.org/abs/2606.13468) Anthropic’s broader internal figures provide context without answering the benchmark question. The company said more than 80% of the code merged into its codebase was authored by Claude as of May 2026, up from low single digits before Claude Code’s February 2025 research preview. It also said the typical engineer was merging eight times as much code per day in the second quarter of 2026 as in 2024, while warning that lines of code overstate true productivity gains. Those figures describe authorship and output, not the acceptance rate of these maintenance routines. [[3]](https://www.anthropic.com/institute/recursive-self-improvement) ## Companies mentioned ## Further sources [[1] Boris Cherny’s LinkedIn post describing Anthropic’s daily Claude maintenance ro… ↗](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7493781303109038080/) [[2] Simon Willison’s edited transcript of a July 2026 fireside chat with Anthropic’… ↗](https://simonwillison.net/2026/Jul/21/cat-and-thariq/) [[3] Anthropic, “When AI builds itself,” describing internal Claude code authorship … ↗](https://www.anthropic.com/institute/recursive-self-improvement) [[4] Empirical study comparing Claude Code pull requests with human-authored pull re… ↗](https://arxiv.org/abs/2509.14745) [[5] Study of rejection reasons for agent-generated pull requests using the AIDev da… ↗](https://arxiv.org/abs/2606.13468) The stories that matter, in one email. Free — unsubscribe anytime.