OpenAI 让 Codex 智能体自动排查 ChatGPT 减速故障:日均 17 次调查,70% 诊断准确
OpenAI 自己在用 Codex 智能体排查 ChatGPT 减速,每天约 17 次、七成诊断靠谱,但 600 个问题里只有约 15 个修复直接上线。想搞清智能体现在到底能干到哪一步的工程团队,这组真实数据比演示视频有用得多。
OpenAI 搭建了一套自动告警排查系统:ChatGPT 一旦出现性能下降,系统会自动启动 Codex 智能体,翻查日志和最近的代码变更来定位原因。流水线由三个环节组成——第一个智能体负责提出诊断,第二个智能体的任务是把诊断证伪,第三个提出修复代码,最终由人类工程师审批合入。
实际效果数据相当具体:这套系统每天大约跑 17 次调查,其中 70% 的诊断被判定为准确。但修复环节明显偏弱——在 600 多个已确诊的问题里,智能体自己写的修复代码原样上线(不经人工改动)的只有大约 15 个。
这说明当前编码智能体在“读日志找根因”这类分析型任务上已经可用,但在独立产出可直接上线的补丁上仍有明显差距,诊断与修复的能力落差对评估智能体工程价值是个有用的参照。
OpenAI built a system where alerts about ChatGPT slowing down automatically start Codex agents, which dig through logs and recent code changes to find the cause. A second agent then tries to prove that diagnosis wrong, and another proposes a code fix that a human engineer approves. It finds causes well, about 17 investigations a day with 70% judged accurate, but fixing is the weak part: out of 600-plus diagnosed problems, only about 15 of the agents’ own fixes shipped unchanged.
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Codex, OpenAI, agents, AI engineering.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai). Cross-source corroboration: 11 independent sources have covered this development (including @OpenAIDevs, IT之家, Windsurf), confirming industry-wide signal strength.