OpenAI Codex Agents Complete More Than 1,700 ChatGPT Investigations
The Performance Factory averages 17 investigations a day. Engineers retain final approval over changes that ship.

OpenAI’s Performance Factory completed more than 1,700 investigations into ChatGPT performance problems in a little over three months, using Codex agents to handle recurring engineering work while leaving final approval with human engineers.
The system averages 17 investigations a day. Agents identified root causes in more than 600 cases, and about 400 of those led to proposed fixes.
Fifteen Codex-authored pull requests went in as written. Agent suggestions also influenced roughly 50 other fixes that engineers completed. About 25% of the diagnoses receive adversarial peer review, in which other agents challenge the findings.
Alerts trigger cloud-based sessions that coordinate several stages: alert management, forensic investigation, peer review and checks to determine whether fixes worked after deployment.
Engineers retain final approval over changes that ship.