In an engineering account published November 19, OpenAI described Performance Factory, an internal system that uses Codex agents to investigate ChatGPT performance regressions, propose fixes and check recovery after deployment. Engineers still review proposed changes and decide whether to merge them, according to OpenAI’s account. Performance Factory is an engineering workflow, not a feature ChatGPT users can turn on.

NeoTeo previously covered Codex in OpenAI’s incident response; Performance Factory is a separate system focused on investigating performance problems.

How the Regression Workflow handles a slowdown

A configured performance alert starts a tracked task. Codex then investigates when a slowdown began, which requests were affected and whether a deployment or configuration change could explain it. OpenAI says its perf-forensics skill guides the agent through that work, using internal code repositories and observability tools such as metrics, logs, traces, CPU profiles and memory snapshots. The agents operate with scoped permissions tied to service identities.

The system coordinates that work through a task service that stores tasks and progress, a dispatch loop that assigns ready tasks, and cloud Codex execution. Subscriptions can resume waiting work after events such as review comments or deployments. In practice, that means an investigation can pause for an engineer’s input or a rollout, then continue when the relevant event arrives.

Why a second Codex session challenges the diagnosis

After an initial investigation, a separate Codex session starts with fresh context to check the supporting evidence and test other explanations. OpenAI says this adversarial review can send a task back for more investigation when the diagnosis does not explain the observed symptoms.

One example in OpenAI’s account involved image requests that continued failing after an internal file-location service had recovered. The review reopened the investigation rather than treating recovery of that service as a complete explanation. That extra challenge matters because identifying a likely cause is not the same as showing that it accounts for every affected request.

Once Codex proposes a fix, an engineer reviews the proposed change and decides whether it should be merged. After deployment, an agent checks whether performance has returned to baseline. If the regression persists or improves only partly, the workflow can send the issue back for further investigation.

Agents can gather evidence and propose a repair; the engineer remains responsible for deciding whether the change ships.

What OpenAI’s reported figures count

OpenAI says the Regression Workflow tracked more than 1,700 investigations and identified root causes in over 600 cases during a period of just over three months. The company gives an average of 17 automated investigations per day. In a separate manual review, OpenAI graded 70% of 100 root-cause suggestions as accurate; the other 30% needed refinement or did not fully establish causality.

MeasureOpenAI-reported figureScope
Investigations trackedMore than 1,700Just over three months
Cases with root causes identifiedOver 600The same reported period
Average investigation volume17 per dayOpenAI’s stated average
Suggestions graded accurate70% of 100OpenAI’s manual review of root-cause suggestions

These figures describe investigation volume and the company’s assessment of suggested diagnoses. They show the scale of the workflow OpenAI reported, while the review result also indicates that some suggestions needed more work before establishing causality.

Where engineering judgment still matters

OpenAI says Codex is a better fit for problems with clear causes and contained changes than for architectural changes. A broad redesign can lead an agent toward an unnecessarily complex fix; for example, caching may introduce stale data. The engineer’s review is therefore a meaningful decision point, not just a final formality.

OpenAI described proactive detection across releases, synthetic performance tests and possible deployment of some low-risk fixes without human approval as work in progress or exploration. In the workflow it described, engineers retain the merge decision.