AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

OpenAI: Codex agents ran more than 1,700 performance investigations in just over three months

AI-generated contentWritten by AI

openai.com, x.com · 2 sources · Sources published: 19 November 2026

OpenAI said agents in its Performance Factory system found the root cause in more than 600 cases. An engineer still decides whether to merge a fix, and 70 of the 100 root-cause suggestions the company reviewed manually were accurate.

OpenAI described Performance Factory in a blog post on 19 November. In this system, Codex agents investigate why the company’s services slow down or become less available. When a performance alert fires, the system opens a task. An agent looks for the cause and turns its fix into a pull request. An engineer reviews it and decides whether to merge it. After deployment, another agent checks whether performance has recovered.

The core idea is that the work does not depend on any single agent’s memory. A task service stores each task’s state and findings. A scheduled loop hands agents the tasks that are ready for their next step. During review or deployment, a task waits, for example for a GitHub comment, so the agent does not spin idly. The agents run in the cloud under a service identity, so their permissions can be scoped and their actions identified. A skill, an instruction package called perf-forensics, guides each investigation. It tells Codex which of the performance team’s tools to use and when.

OpenAI said the system’s first workflow, the Regression Workflow, logged more than 1,700 investigations in just over three months and found the root cause in more than 600 cases. Finding the cause does not mean the fault has been fixed, however. When the company manually reviewed 100 of the agents’ root-cause suggestions, it graded 70 as accurate. The rest needed refinement or did not fully establish causality. All the figures are the company’s own.

At first, the agents sometimes settled for the first plausible explanation, and some diagnoses turned out to be wrong. A separate Codex session therefore now tries to disprove every diagnosis. In one case, an agent blamed image loading for errors in ChatGPT conversations. The reviewing session noticed, however, that image requests were still failing even though the service that locates files had already recovered. It reopened the investigation.

Writing fixes has been harder than finding the cause, OpenAI said. Some of the agents’ proposals have been needlessly complex or introduced new problems, so they go through a similar review. Some fixes require an architectural decision: a cache can speed up a service but serve stale data. The company is exploring whether agents could in future deploy low-risk fixes without a human. For now, an engineer makes the decision.

Report an error in this story