OpenAI says Codex-powered Performance Factory found root causes in over 600 cases; engineers decide each merge

By
CTOL Staff Reporter
1 min read

Codex agents run roughly 17 performance investigations a day, but engineer approval still separates diagnosis from production repair.

At OpenAI, a Codex agent can trace a slowdown through performance evidence and propose a code change. The last step is strikingly human: an engineer decides whether the pull request gets merged. On 19 November 2026, OpenAI described its internal Performance Factory as an autonomous system for investigating performance regressions. The agents can hunt for causes. People still choose which fixes reach production.

We would credit that division of labor as a real operating advantage, while refusing to value it as autonomous production repair. The Diagnosis Deployment Split is the separation between an agent finding and proposing a fix and an engineer authorizing a change to live software. OpenAI's numbers describe the first activity. Its approval rule governs the second. Faster diagnosis can free engineering attention, but a persuasive answer and a safe production change are different deliverables.

According to OpenAI's 19 November 2026 engineering account, the Factory tracked more than 1,700 investigations and found root causes in over 600 cases in just over three months. It handled around 17 automated investigations a day. Taking the two rounded lower figures, 600 divided by 1,700 gives about 35.3% root cause coverage, an approximation because both counts exceed those figures. That is a considerable stream of investigative work, even when an explanation is hard to find. It says something narrower about the financial return from fixes that actually make a service faster.

OpenAI also manually graded 100 root cause suggestions and called 70% accurate. The other 30% needed refinement or failed to establish causality. That 70% grade covers the sample of 100 suggestions only, rather than all 600 cases with identified root causes. Accuracy within a sample and coverage of a larger investigation stream measure different things. Neither is a count of approved repairs. It would be easy to celebrate the 600 as completed fixes, especially when the Factory generates proposed pull requests. A diagnosis saves investigative effort; the reliability benefit follows when an engineer accepts a safe improvement and the service recovers.

The agents run inside OpenAI's network, using a service identity with repository and observability access. Engineers receive summaries, time series, flame graphs and traces, while related incidents are grouped and ranked by estimated impact. OpenAI had accumulated more than 30 performance tools before adding Codex. The preparation gives engineers evidence and an indication of priority. The engineer can allocate attention to problems worth solving, instead of starting every search from scratch. That is a credible source of operating leverage.

Early agents sometimes stopped at the first plausible explanation, and some were wrong. The system added an adversarial review session. Proposed fixes sometimes grew more complicated than necessary or introduced new issues, and OpenAI says getting fixes right has been harder than finding causes. Its engineering account explicitly identifies review and merging as today's bottleneck. When an agent produces diagnoses more readily, the scarce skill shifts toward deciding which proposed change is safe. A faster queue of explanations can still end at a slow queue of approvals.

The economic exposure starts with the importance of a working service. In an announcement on 31 August 2026, OpenAI put its advertising business at a $1 billion annualized revenue run rate and said its service was available to more than 1 billion weekly active users. Take a scenario in which the service stalls completely for one hour. Dividing that $1 billion run rate by 365 days and 24 hours gives about $114,000 of advertising run rate exposed during the hour, before subscriptions or API revenue. That is scenario exposure, rather than an observed outage loss. An unsafe repair can erase the value of many correct diagnoses.

The opposing case is stronger than the usual argument that agents will someday write better code. During the same just over three months, around 17 investigations a day ran automatically, and more than 600 reached a root cause. If the expensive work is gathering evidence and narrowing possibilities, human approval is a sensible safety choice rather than an automation failure. Engineers can spend their attention on judgment instead of conducting every investigation. On that definition, the Factory already does economically useful autonomous work. That same period supplies the limit: over 600 of more than 1,700 investigations reached a root cause, and every proposed fix still went to an engineer for the merge decision. It saves investigative effort while leaving production risk with engineers.

Microsoft provides an encouraging benchmark for diagnosis. Its RCACopilot paper dated 13 November 2023 says, "RCACopilot achieves RCA accuracy up to 0.766." The result draws on a year of real incidents. Meta's 24 June 2024 engineering account recorded 42% backtested accuracy at identifying root causes for web monorepo investigations when they were created, using a top five suggestion measure. These measures describe distinct diagnostic tasks, so neither is directly comparable to OpenAI's approximate 35.3% coverage. The Factory tracked more than 1,700 investigations in just over three months, while engineers retained authority to merge each proposed fix.

Google DeepMind shows what the opposite outcome can look like. In its 17 August 2018 account, it described moving from operators manually implementing AI cooling recommendations to direct AI control under operator supervision. Average energy savings were around 30%, and operators could exit AI control mode. That change removed the routine human implementation step while preserving an override. It makes a persuasive case for granting a system more authority once the task is bounded and feedback is immediate. Production code changes have less predictable consequences than cooling adjustments. The precedent argues for pursuing independent repair, and for demanding much stronger safeguards before the final merge decision moves to an agent.

OpenAI says it is exploring low risk fixes that could ship without a human in the loop. The Diagnosis Deployment Split would narrow, because the machine would own a portion of the last consequential decision. More accurate root cause suggestions remain valuable, but the critical next improvement is a change process that engineers can safely delegate. We would want to see permission to ship tied to actual production behavior, rather than a larger stack of plausible pull requests. OpenAI already verifies recovery after deployment; removing human approval would make that verification, and the ability to reverse a mistake, still more important.

We would give OpenAI credit for automating the hunt for performance regressions, but assign no premium for autonomous production repair while engineers decide which fixes ship. We would be wrong if, by 19 May 2027, a public OpenAI engineering update confirms that Performance Factory fixes ship without engineer review.

You May Also Like

This article is submitted by our user under the News Submission Rules and Guidelines. The cover photo is computer generated art for illustrative purposes only; not indicative of factual content. If you believe this article infringes upon copyright rights, please do not hesitate to report it by sending an email to us. Your vigilance and cooperation are invaluable in helping us maintain a respectful and legally compliant community.

Subscribe to our Newsletter

Get the latest in enterprise business and tech with exclusive peeks at our new offerings

We use cookies on our website to enable certain functions, to provide more relevant information to you and to optimize your experience on our website. Further information can be found in our Privacy Policy and our Terms of Service . Mandatory information can be found in the legal notice