AI News AI资讯 7d ago Updated 7d ago 更新于 7天前 48

Claude Code now runs daily maintenance on Anthropic's software with a 46 percent merge rate Claude Code 现在每天对 Anthropic 的软件进行维护,合并率达 46%

Anthropic is testing Claude Code for autonomous daily maintenance of its own software, with Boris Cherny overseeing the experiment Over a few weeks, Claude generated 388 pull requests across iOS, Android, desktop, web, CLI, and Agent SDK platforms, achieving a 46% merge rate after human review Twelve specialized maintenance routines were deployed, including Crash Fuzzer, Dup Unifier, Dead Code Removal, Flaky-Test Fixer, and Logic Bug Fixer The system operates through a dedicated Slack channel us Anthropic工程师Boris Cherny让Claude Code自动执行内部软件日常维护,几周内生成了388个PR,其中180个被合并,合并率达46% Claude通过12种专用维护例行程序运行,包括崩溃模糊测试、重复代码合并、死代码删除、逻辑简化、不稳定测试修复等 维护流程通过Slack频道"proj-claude-maintains-apps"调度,覆盖iOS、Android、桌面、Web、CLI和Agent SDK全平台 提示工程极为简单,用自然语言直接描述任务即可,无需复杂提示设计 当前超过一半PR未通过审核,作者称这是"自主AI应用维护的早期信号",Anthropic正探索加

72
Hot 热度
65
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Anthropic is testing Claude Code for autonomous daily maintenance of its own software, with Boris Cherny overseeing the experiment
  • Over a few weeks, Claude generated 388 pull requests across iOS, Android, desktop, web, CLI, and Agent SDK platforms, achieving a 46% merge rate after human review
  • Twelve specialized maintenance routines were deployed, including Crash Fuzzer, Dup Unifier, Dead Code Removal, Flaky-Test Fixer, and Logic Bug Fixer
  • The system operates through a dedicated Slack channel using plain-language prompts without elaborate prompt engineering
  • Cherny describes results as "early signs of life" for autonomous AI-powered app maintenance, with ongoing efforts to improve merge rates

Why It Matters

This experiment represents one of the most concrete real-world demonstrations of AI agents handling sustained, autonomous software maintenance at scale within a major tech company. For AI practitioners and engineering leaders, it provides empirical data on what current LLM-based coding agents can and cannot handle in production environments, offering a realistic benchmark for autonomous developer tooling.

Technical Details

  • Architecture: Claude Code operates via Tag integration, running daily routines through a dedicated Slack channel ("proj-claude-maintains-apps") across all Anthropic platforms
  • Twelve specialized routines: Crash Fuzzer (simulator-based crash detection and fix generation), Dup Unifier (merges similar abstractions), Dead Code Removal (with logging verification), Logic Simplifier, Logic Bug Fixer, Useless Test Pruner, Flaky-Test Fixer, Abstraction Improver, Abstraction Police, Ant-only Shipper, and Shipped-Feature Inliner
  • Prompting approach: No elaborate prompt engineering; Cherny uses plain-language Slack messages instructing Claude to run routines on real apps without mocks
  • Performance metrics: 388 PRs generated, 180 merged (46% rate); Claude typically gets PRs right on first attempt, with iterative tuning taking a few days when adjustments are needed
  • Review pipeline: Combination of automated Claude Code review and human review before merging

Industry Insight

  • The 46% merge rate suggests autonomous AI maintenance is viable for mechanical, well-scoped tasks but still requires significant human oversight for complex or high-risk changes; companies should start with narrow, low-stakes maintenance routines before expanding scope
  • The plain-language prompting approach demonstrates that sophisticated prompt engineering may be less critical than previously assumed for agent-based development workflows, potentially lowering the barrier to adoption
  • The iterative tuning model—where routines are refined over days based on failures—indicates that organizations should expect a learning curve and build feedback loops into their AI agent deployment strategies rather than expecting plug-and-play reliability

TL;DR

  • Anthropic工程师Boris Cherny让Claude Code自动执行内部软件日常维护,几周内生成了388个PR,其中180个被合并,合并率达46%
  • Claude通过12种专用维护例行程序运行,包括崩溃模糊测试、重复代码合并、死代码删除、逻辑简化、不稳定测试修复等
  • 维护流程通过Slack频道"proj-claude-maintains-apps"调度,覆盖iOS、Android、桌面、Web、CLI和Agent SDK全平台
  • 提示工程极为简单,用自然语言直接描述任务即可,无需复杂提示设计
  • 当前超过一半PR未通过审核,作者称这是"自主AI应用维护的早期信号",Anthropic正探索加速机械性变更的合并流程

为什么值得看

本文展示了Claude Code在真实生产环境中的自主代码维护能力,为AI辅助软件工程提供了可量化的实践数据。对AI从业者和开发者而言,这验证了大模型在代码理解、修复和重构任务上的实用潜力,同时也揭示了当前AI自主维护的局限性与改进方向。

技术解析

  • 维护架构:Claude Code通过Slack频道接收调度指令,每日自动执行12类代码维护例行程序,覆盖Anthropic全平台(iOS、Android、桌面、Web、CLI、Agent SDK)。
  • 核心例行程序:包括Crash Fuzzer(在模拟器中随机触发崩溃并生成修复)、Dup Unifier(合并相似抽象)、Dead-Code Remover(删除不可达代码,可疑代码先加日志次日验证)、Logic Simplifier/Bug Fixer、Flaky-Test Fixer、Shipped-Feature Inliner(移除已上线功能标志)等。
  • 提示设计:无需复杂prompt工程,用自然语言直接描述任务(如"对iOS/Android/桌面启动崩溃模糊测试,使用真实应用而非mock,触发崩溃后生成PR")。
  • 效果数据:几周生成388个PR,经自动审查+人工审核后合并180个,合并率约46%;多数PR首次提交即正确,未通过的主要是机械性变更,Anthropic正优化合并流程。

行业启示

  • AI代码维护进入实用阶段:46%的合并率表明AI已能可靠处理大量重复性、机械性代码维护任务,可显著释放开发者人力。
  • 人机协作模式仍需人工审核:超半数PR未通过说明当前AI在复杂逻辑判断和架构理解上仍有局限,"AI生成+人工审核"仍是必要模式。
  • 提示简单化降低使用门槛:无需复杂prompt工程即可驱动AI执行多步骤维护任务,意味着企业可快速将Claude Code等工具集成到现有开发流程中。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Claude Claude Code Generation 代码生成 Agent Agent Programming 编程 Product Launch 产品发布