Skip to content
Advertisement
OpenAI’s Performance Factory uses Codex agents with adversarial peer review

OpenAI’s Performance Factory uses Codex agents with adversarial peer review

OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment

An internal system of Codex agents runs about 17 performance investigations a day, while human engineers keep final approval

OpenAI has built an internal system that hands one of software engineering’s least glamorous jobs to AI agents. It is called the Performance Factory, and it uses Codex coding agents to hunt down performance problems in ChatGPT.

The system runs an average of 17 automated performance investigations each day.

OpenAI announced the system on November 19, 2026, through a report by VentureBeat and a thread on its @OpenAIDevs account. The pitch is simple. Let agents handle the tedious forensics, and let humans make the final call.

What the factory actually does

In software, a performance regression happens when a change makes something slower or less reliable than it was before.

Traditionally, that someone is an engineer digging through logs by hand. OpenAI says the Performance Factory replaces much of that manual log analysis with an automated workflow. Alerts trigger it, and the work runs in cloud-based sessions.

The system coordinates several agents across distinct stages. These include managing alerts, running detailed forensic investigations, conducting peer reviews, and checking whether fixes actually worked after deployment.

The numbers so far

In a little over three months, the Performance Factory has completed more than 1,700 investigations. Those 17 daily runs add up fast.

Advertisement

Of those investigations, the agents identified root causes in more than 600 cases.

From that pool of root-cause cases, around 400 led to proposed fixes.

The share of fixes written entirely by the machines is much smaller. Just 15 were exact Codex-authored pull requests, meaning code changes the agents drafted that went in as written.

Agent suggestions influenced roughly another 50 fixes. In those cases, the Codex output pointed engineers in the right direction even if humans did the final writing.

Then there is the quality control layer. Approximately 25% of the diagnoses Codex agents produce get challenged through an adversarial peer review process.

Why OpenAI built it this way

The design reflects a broader shift in how OpenAI uses Codex. Instead of asking the agents to generate a single chunk of code on request, the company is chaining them into integrated workflows that tackle recurring operational headaches.

OpenAI is also explicit that humans stay in the loop. Engineers keep final approval over what gets shipped. The agents investigate and propose, but they do not get the keys to the building.

That balance shows up in the data. The gap between roughly 400 proposed fixes and 15 exact Codex-authored pull requests suggests the system is mostly accelerating human work rather than replacing it.

What this means for engineering teams

More than 1,700 investigations in a little over three months is volume that would strain a human team, especially for alerts that may turn out to be nothing.

The funnel narrows: investigations narrow to root causes, root causes narrow to proposed fixes, and only a sliver become pure agent-written code.

The adversarial review step, with approximately 25% of diagnoses getting pushback, is an admission that agent output still needs checking before anyone acts on it.

The post-deployment verification stage exists to catch confident but wrong diagnoses that can waste time just as easily as they save it.

OpenAI is effectively using its own product to run its own product. Every investigation doubles as a stress test of Codex in a demanding, real-world setting.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.