OpenAI's Performance Factory Uses Codex Agents to Investigate ChatGPT Performance Issues
Key Takeaways
- •OpenAI's Performance Factory uses Codex coding agents to automate performance regression investigations in ChatGPT, averaging 17 automated investigations per day.
- •The multi-agent workflow spans alert management, forensic investigation, adversarial peer review, and post-deployment verification, with human engineers keeping final approval over what gets shipped.
- •In a little over three months, the system completed more than 1,700 investigations, identified root causes in over 600 cases, and led to roughly 400 proposed fixes.
- •Only 15 fixes were exact Codex-authored pull requests, with about 50 more influenced by agent suggestions, indicating the system mostly speeds up human engineering work rather than replacing it.
- •Approximately 25% of Codex agent diagnoses are challenged through an adversarial peer review process, serving as a built-in quality-control layer before fixes are acted upon.

OpenAI has built an internal system that hands one of software engineering's least glamorous jobs to AI agents. The system, known as the Performance Factory, uses Codex coding agents to track down performance problems in ChatGPT, running an average of 17 automated performance investigations per day while human engineers retain final approval over what gets shipped.
OpenAI announced the system on November 19, 2026, through a report by VentureBeat and a thread on its @OpenAIDevs account on X. The pitch is straightforward: let agents handle the tedious forensic work, and let humans make the final call.
What the Factory Does
In software, a performance regression occurs when a change makes a system slower or less reliable than it was before. Traditionally, identifying one means 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 the process, and the work runs in cloud-based sessions.
The system coordinates several agents across distinct stages, including managing alerts, running detailed forensic investigations, conducting peer reviews, and verifying 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. Of those, the agents identified root causes in more than 600 cases. Around 400 of those root-cause cases led to proposed fixes.
The share of fixes written entirely by 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, with Codex output pointing engineers in the right direction even though humans did the final writing.
There is also a 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. Rather than asking 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 a volume that would strain a human team, particularly for alerts that may turn out to be nothing. The funnel narrows at each stage: 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 acknowledgment that agent output still needs checking before anyone acts on it. The post-deployment verification stage exists to catch confident but wrong diagnoses, which can waste time 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.
Those published figures also give teams outside OpenAI a concrete baseline for evaluating agent-assisted engineering claims, and the number worth watching is the share of proposed fixes that land as exact agent-authored pull requests — a directly reported measure of where the boundary between agent and human work currently sits in a production system.
This article is based on reporting by CryptoBriefing, edited by Diego Almada Lopez.