OpenAI details how Codex investigates ChatGPT slowdowns

OpenAI described its Performance Factory on November 19, an internal system that uses Codex to investigate ChatGPT slowdowns, suggest fixes and check whether performance recovers. Engineers decide which fixes ship.
When an alert fires, the system opens a task and assigns an agent to look for the cause. A separate Codex session challenges that diagnosis. The task holds the findings outside the agent’s conversation, so an engineer can review the suggested change and another agent can pick up the same task after deployment. Subscriptions can wake an agent when a code review comment arrives or a deployment finishes.
OpenAI says the workflow tracked more than 1,700 investigations and identified root causes in over 600 cases in just over three months. In a separate manual review of 100 suggested causes, it graded 70% as accurate. The rest required refinement or did not establish causality.
The system shows how an engineering team can delegate repeatable incident work without handing over the decision to change production software. OpenAI says it is exploring how agents might eventually ship low-risk fixes.
Sources and citations
Revision history
Published .