The News, Explained

OpenAI introduced Performance Factory on November 19, 2026, describing an internal system for investigating and fixing performance problems in services including ChatGPT. Codex, its coding agent, investigates and proposes changes, while engineers review the proposed code and decide whether to merge it. The announcement gives users a view of the decisions and checks behind responding to a slowdown. Source

The first workflow handles performance regressions: cases where a software change leaves response times or availability worse than before. The source gives ChatGPT’s first-token latency as one investigation target. Tokens are pieces of text making up an answer, and first-token latency is the wait until the first piece appears. An alert creates a task, and an agent investigates when performance deteriorated and which requests were affected. Source

A slower average still leaves questions about the cause. The agent checks whether comparable requests became slower or whether the service received more requests that were already slow to process. It then examines deployments, configuration changes, code and execution records. Distinguishing a change in the mix of requests from the effect of a particular change helps identify what needs fixing. Source

A separate Codex session challenges the first diagnosis, looking for evidence that could disprove its assumptions and comparisons. In the company’s image-download example, image requests continued failing after the internal service blamed for the problem had recovered. The reviewing agent reopened the investigation because that explanation did not account for the continuing failures. Source

OpenAI says the workflow tracked more than 1,700 investigations and identified root causes in over 600 cases in a little over three months. In a separate human review of 100 root-cause suggestions, it graded 70% as accurate; 30% needed refinement or had not sufficiently established causality. Investigation counts, cases with a root cause and accuracy within this sample measure different stages and populations. Source

Once the diagnosis passes review, Codex implements a proposed fix and opens a code-change request. Engineers inspect the evidence and proposal and decide whether to merge it; they can also edit the code or write their own fix from the diagnosis. OpenAI reports proposals that were unnecessarily complicated or introduced new problems. A design such as a cache can trade faster processing for the risk of stale data, benefiting from an engineer who understands the service. Source

Work continues after deployment. Another agent checks whether performance returned to its earlier baseline. Partial recovery opens a new task for the remaining slowdown; no improvement sends the recorded result back for investigation. The system preserves task states and evidence outside any individual agent’s conversation, then resumes work after events such as review comments or completed deployments. Source

The company says it evaluates investigation quality separately from the usefulness of summaries passed to engineers when changing models or instructions. Tests based on past incidents also need checks that later evidence has not given away the answer. Shipping low-risk fixes without a human and proactively detecting slowdowns that alerts miss are directions being explored or developed. Source

OYOPICK’s Take

We put more weight on the connection between deployment, recovery checks and reopened investigations than on the number of fixes an AI proposes. Read alongside the company’s finding that 30% of 100 root-cause suggestions needed refinement or had not established causality, that connection makes a case for expanding the checks that revisit conclusions as automated investigation grows.

The benefit for everyday users should become easier to assess as evidence of service recovery and recurrence accumulates. We would welcome future reporting that shows recovery and remaining problems within a consistent scope. If the evidence for human decisions and the link back to actual outcomes become clearer, readers could better assess the prospect of agents and engineers sharing recurring operational work.