OpenAI now has Codex agents investigating ChatGPT slowdowns on their own, then handing a proposed fix to a human to approve. The company described the system, called the Performance Factory, in an engineering post published November 19 by Okay Zed and Cyrus Forbes, both members of technical staff. Its first workflow handled more than 1,700 investigations in a little over three months, according to OpenAI, and traced over 600 of them to a root cause. That works out to about 17 investigations a day.

The arithmetic deserves a second look. If 600 is the root-caused count out of 1,700, roughly a third of cases ended with a named cause. OpenAI does not say what happened to the rest, how many fixes engineers merged, or how much latency came back as a result.

The mechanics are plain. An engineer writes a workflow as a configuration file plus Markdown prompts, listing the states a task can occupy and which steps need a person. A dispatch loop checks which tasks are ready and starts a Codex turn for each one, running inside OpenAI’s network under a service identity, so permissions can be scoped and agent actions audited. When a task must wait for a code review or a deployment, it subscribes to that event and sleeps instead of burning an agent session.

The Regression Workflow, the first one built, opens a task when a performance alert fires. An agent works out when the metric started slipping, which requests were hit, and whether a deployment or feature flag lines up with the timing. A later agent drafts a fix. After deployment, another checks whether performance returned to baseline, and reopens the case if it did not.

Accuracy is the weak spot, and OpenAI says so. The team hand-graded 100 root-cause suggestions and called 70 of them accurate. The other 30 needed more work or did not fully establish causality. Early agents tended to stop at the first plausible explanation, so OpenAI added a second Codex session whose job is to disprove the first one’s diagnosis.

The post gives one example. An agent blamed failing ChatGPT conversations on image download errors. The reviewing session saw that the file-lookup service was healthy again even though image requests kept failing, so the stated cause could not account for the symptom, and it reopened the case. That is a model checking a model, and the 70 percent figure comes from OpenAI’s own grading, not an outside audit.

Fixes have been harder than diagnoses. OpenAI says agent-written patches were sometimes more complicated than needed or introduced new problems, so each proposed fix now gets its own adversarial review. Some changes also need judgment no agent has. A cache can speed a service up while serving stale data, and an engineer who knows the service has to weigh that trade.

The most useful detail for anyone building something similar is how OpenAI spent its effort. It reports that raw agent output was a wall of text, so the team now ties related incidents together, sorts them under the metrics teams track, and ranks each by estimated impact. It tests changes to the agents themselves by replaying past incidents with the answer stripped out, and runs a daily agent that reads recent logs for broken tools and missing permissions.

OpenAI says it is exploring agents that scan profiles and traces for regressions no alert catches, and agents that ship low-risk fixes without a human reviewer. Teams running their own incident agents should copy the 100-case grading exercise first: measure how often your agent’s diagnosis is wrong before deciding how much review you can afford to remove.

OpenAI, “How OpenAI’s Performance Factory keeps ChatGPT fast,” engineering post by Okay Zed and Cyrus Forbes, November 19, 2026.