How the factory runs

The path to the system is instructive because it is mostly plumbing, not a new model. OpenAI's performance team started with manual investigations and more than 30 specialized tools. Codex then took over the searching: an engineer could ask why a service had slowed and get an answer assembled from logs and telemetry. The team captured its own procedure in a reusable Codex skill called perf-forensics, so an alert could produce a diagnosis in a single run. At that point the engineers had become couriers, copying alerts into Codex and pasting findings into Slack, so the team connected the alerts directly and gave each regression a task record that agents advance through stages: investigation, root cause, code change, review, verification.

The task record is the design decision worth copying. Agents store high-level findings, the metrics that matter and the queries that produced them, rather than raw profiling dumps. The next agent reads a compact record and can rerun the queries if it wants to check the evidence itself. Four components move the work: the task service, a dispatch loop that starts a Codex session when a task is ready, a cloud API for running those sessions against OpenAI's repositories and observability stack, and subscriptions that let a task sleep until, say, an engineer comments on a pull request.

The second agent's job is to disagree

OpenAI found that its investigator agents stopped at the first plausible explanation, even when the evidence did not fully support it. The response was structural rather than a sterner prompt: a separate Codex session, starting from fresh context with the full task record, is instructed to find gaps in the diagnosis, verify the evidence and test alternative explanations.

The example OpenAI gives shows the value of the step. In one investigation, the first agent blamed failing ChatGPT conversations on errors in image downloads. The reviewer challenged that explanation because the internal service responsible for locating files had already recovered while image requests kept failing, so the diagnosis could not explain the continuing failures, and the case went back for more work. Early on, the reviewer returned a substantial share of diagnoses; the team fed that criticism back into the investigator's instructions, and the return rate now sits at around 25%. That return rate is a workload figure, not an error rate: it counts diagnoses sent back for more work, not diagnoses proven wrong. OpenAI's blog post, as reported in summaries of it, also says the team graded 100 root-cause suggestions by hand and rated 70% accurate.

Read the funnel, not the headline count

The honest part of this disclosure is the arithmetic at the end, and OpenAI states it plainly.

More than 600 investigations reached a root cause. Around 400 of those produced a proposed fix. Approximately 50 fixes that eventually shipped were influenced by or matched what an agent suggested. Around 15 shipped fixes were the exact pull requests Codex authored. The reporting does not establish that the other proposals were rejected; some may still be pending. OpenAI itself says these results fall short of its ambitions for automated remediation, and that fixes are where the agents are weakest: they perform best when the cause is clear and the change is contained, like spotting that a request serializes the same conversation several times and proposing it be done once. On feature work that adds latency and needs architectural change, agents propose fixes that are too complicated or introduce new problems, which is why proposed fixes now get their own adversarial review.

That funnel is the useful takeaway for any team running coding agents today. The layer the system automates, on this evidence, is the investigation: detection, triage, evidence gathering, tracking and verifying recovery. OpenAI says automating that layer is what freed its engineers for longer-term performance work. The layer that still consumes human judgment is the change itself, and OpenAI has not closed that gap; it is exploring whether low-risk fixes could one day ship without a human, which is an admission that today they do not.

One caution applies to the whole account. These are OpenAI's own numbers about its own internal system, shared through an exclusive report; there is no independent measurement, and the coverage gives no measured improvement in end-user latency. A system that runs 17 investigations a day is real operational evidence. It is not yet evidence about how much faster ChatGPT actually got.

Sources

  1. [1] VentureBeat, "Exclusive: OpenAI reveals Performance Factory, a Codex-powered software optimization system that keeps ChatGPT fast," November 19, 2026Read source
  2. [2] OpenAI, "How OpenAI's Performance Factory keeps ChatGPT fast," blog post by Okay Zed and Cyrus Forbes, November 19, 2026 (announced via @OpenAIDevs; the 70%-of-100 hand-grading figure is OpenAI's own, as reported in summaries of the post)
  3. [3] AICoder live log, "OpenAI Reveals Performance Factory," November 19, 2026 (70%-of-100 grading figure, attributed to OpenAI's post)Read source