What A Time To Be Alive
Choose language: English

OpenAI uses Codex agents to investigate performance problems in ChatGPT

OpenAI lets Codex agents work out for themselves why parts of ChatGPT slow down. In just over three months they found the cause in more than 600 cases. An engineer still decides whether a proposed fix goes in.

Transparent glass tube with a stream of white light that narrows in the middle; at that point the light turns orange and two light-blue rings circle the tube. In the lower right, out of focus, a hand rests next to a green button.
Illustration: AI agents find out where the software gets stuck, and a human decides on the fix.

If part of ChatGPT slows down after a change to the software, someone has to find out why. At OpenAI, AI agents do that more and more often by themselves. the company described its internal system for this, the Performance Factory. It uses agents from Codex, OpenAI's programming assistant. OpenAI does not say which model version is behind it.

From alert to check

The first workflow focuses on slowdowns that appear after a change: something used to work faster than it does now. When a performance alert goes off, the system creates a task. An agent works out since when things have been slower and which requests are affected. It then searches code changes, measurements and log files, in which the software records what happens, for the cause. In doing so it follows a fixed method that the team had first written out for Codex.

Then a second agent comes in, one that does not know the earlier work. It gets the whole case file and has to look for gaps in the reasoning and test other explanations. It sends back about a quarter of the diagnoses. In one example from OpenAI, the first agent blamed errors when downloading images. The checker saw that the service that looks up the files was already working again while the errors went on, and reopened the investigation.

If a diagnosis holds up, Codex writes a fix and submits it as a proposed code change. An engineer reads the code and decides whether it goes in. After the rollout, an agent checks again whether the problem has really gone. If not, the investigation starts again.

What it has delivered so far

In just over three months, the system tracked more than 1,700 investigations, on average about 17 a day. In more than 600 cases it found the cause, and in about 400 of those Codex proposed a fix. About 50 fixes that actually went into ChatGPT were based on a proposal from the agents or matched one. Around 15 were exactly the code change that Codex had written. According to OpenAI, that is still short of what the team wants to achieve.

The agents do well when the cause is clear and the fix is small, such as preparing a conversation for sending once instead of several times. Complicated fixes are harder, such as a temporary store that keeps frequently requested data but can therefore show outdated data. Meanwhile, engineers do get bundled reports, ranked by how many users are affected, with short summaries and charts. As next steps, OpenAI itself names looking for improvements by comparing versions, and research into whether agents may one day roll out simple, low-risk fixes without a human's approval.

Why OpenAI needs this

Within OpenAI, Codex has become the tool for almost all work, including outside the engineering teams. In September 2026, Venkat Venkataramani, vice president of engineering at OpenAI, told the newsletter The Pragmatic Engineer that the load on some systems for building, testing and rolling out software had grown about tenfold in roughly six months. As a result, far more code passes through those systems than engineers can check by hand.

What does this mean for the road to the singularity?

AI agents that write code on request are already common. What is new is that here agents run the whole maintenance loop: they notice a problem, look for the cause, let a second agent attack the reasoning, propose a fix and check afterwards whether it worked. The human still sits at one point in that loop: the decision whether the code goes in. Fifteen code changes from Codex itself in just over three months is a first step, but the method is in place.

As such agents get better, the work of engineers shifts from searching to judging. That will also be necessary: anyone who produces ten times as much code cannot check it by hand ten times as often. Other software companies that write much more code with AI run into the same limit and will also turn to such checking agents. The most attractive prospect is software that largely keeps itself healthy, so that people spend their time on new things instead of on the same repairs again and again. The question is where to take the human out of the loop; OpenAI is first studying that with small, low-risk fixes. Meanwhile, AI is already maintaining part of the software of an AI service here, a small example of AI speeding up the work on AI itself.

What is the technological singularity? Read the explanation →

Sources

  1. RuntimeWire
  2. newsletter.pragmaticengineer.com
  3. tldr.tech
  4. OpenAI
  5. VentureBeat