原帖引用 · AI 翻译
标题：“在开发者心流状态中及时介入：谷歌规模下的低延迟智能体程序修复”
谷歌的一篇论文表明，AI 智能体可以修复谷歌内部未通过的测试，而修复速度决定开发者是否会采用修复方案。
合并前在 CI（持续集成）中失败的测试会打断开发者的工作状态，而现有修复智能体大多在合并后离线工作。
他们构建了 FlowAgent，它会捕获每次测试失败，并运行一个智能体，让其编辑代码、构建并重新运行测试，直到测试通过。经过验证的修复方案会出现在代码评审中，并附有简短的根因说明。
专家对 195 次真实故障的评估认为，其 67.18% 的修复方案是正确的；开发者在生产环境中应用了 28,554 项修复。其运行时间中位数为 9.85 分钟，而开发者开始编辑代码前的时间中位数为 27.48 分钟，但许多修复方案仍然来得太晚。
原文
Google paper shows that an AI agent can fix failing tests across Google, and speed decides whether developers use the fix. Tests that fail in CI ( continuous integration) before merge pull developers out of their flow, and existing repair agents mostly work offline after merge. They built FlowAgent, which catches each test failure and runs an agent that edits, builds, and reruns tests until they pass. Verified fixes appear in code review with a short note on the root cause. Experts judged 67.18% of its fixes correct on 195 real failures, and developers applied 28,554 fixes in production. Its median run took 9.85 minutes, versus a 27.48-minute median before developers start editing, yet many fixes still arrived too late.