ENGINEERING / LATEST COVERAGE
OpenAI puts Codex to work on ChatGPT slowdowns.
OpenAI describes how Codex helps engineers trace slowdowns, propose fixes and check that performance recovers.
What changed?
When ChatGPT takes longer to start answering, an engineer has to figure out what changed. Did the same kind of request start taking longer, or did more people start sending requests that already take a while? Both can make a dashboard look worse, but they call for different responses.
In a November 19 engineering report, OpenAI describes a system it already uses to help answer questions like that. The company calls it the Performance Factory. Its first workflow uses Codex to investigate certain alerts, suggest repairs for an engineer to review and check whether a deployed fix restored performance.
Follow the clock.
One useful clue is timing. In a case OpenAI describes, an agent investigating failed ChatGPT conversations pointed to trouble downloading images. A second Codex agent noticed that one service blamed in the diagnosis had recovered while image requests were still failing. The two timelines did not fit, so it reopened the investigation.
That distinction is what makes the example useful: the second agent did not have to know the right answer to identify a gap in the first one. The job stays open, and the next person or agent can see which explanation was tested and what still needs to be established.
A proposed fix isn’t a recovery.
When the workflow has a diagnosis and a proposed code change, an engineer can review the evidence, edit the proposal or write a different fix. After an engineer deploys a change, an agent checks production measurements against their earlier baseline. If performance has not recovered, the workflow can send the task back for another look.
That last check matters because a plausible explanation and a tidy patch are not the same as a faster service. The same measurements that made the slowdown visible can help confirm whether it has actually cleared.
For teams putting Super Intelligence (SI) tools to work, there’s a practical lesson here. An agent doesn’t need to own the whole job to make itself useful. It can pick up an alert, leave a clear trail for the engineer making the decision, and come back after deployment to see whether the decision worked.