How OpenAI’s Performance Factory keeps ChatGPT fast
OpenAI describes Codex workflows that carry performance alerts through diagnosis, proposed fixes, engineer review, deployment, and recovery checks. The process shows how persistent task state and event triggers can support production agent work.
OpenAI’s Performance Factory uses a task tracker, dispatch loop, cloud Codex runs, and event subscriptions to move work between agents and engineers. In just over **three months**, its Regression Workflow tracked **1,700+ investigations** and root-caused **600+ cases**. A manual review rated **70%** of 100 root-cause suggestions accurate.
For recurring operational work, model the workflow as explicit states with task findings stored outside any agent turn. Use events such as review comments or deployment completion to resume work, and keep engineers responsible for approving and shipping fixes. Separate turns can investigate, challenge a diagnosis, propose a fix, and verify recovery.
OpenAI’s Performance Factory uses a task tracker, dispatch loop, cloud Codex runs, and event subscriptions to move work between agents and engineers. In just over **three months**, its Regression Workflow tracked **1,700+ investigations** and root-caused **600+ cases**. A manual review rated **70%** of 100 root-cause suggestions accurate. For recurring operational work, model the workflow as explicit states with task findings stored outside any agent turn. Use events such as review comments or deployment completion to resume work, and keep engineers responsible for approving and shipping fixes. Separate turns can investigate, challenge a diagnosis, propose a fix, and verify recovery. The 70% figure indicates diagnosis gaps, and the article says proposed fixes can be unnecessarily complex or introduce issues. Results are from OpenAI’s internal systems; the material does not establish how the approach performs elsewhere.
The Performance Factory account gives operational detail for a recurring agent loop: tracked investigations, event-triggered handoffs, and human approval, alongside a 70% accuracy rating for root-cause suggestions and risks from complex fixes. It reinforces the need for durable, reviewable traces and explicit runtime states, while limiting claims of reliable autonomous repair.