How OpenAI's Performance Factory Uses AI Agents to Keep ChatGPT Fast
💼 Business AI News

How OpenAI's Performance Factory Uses AI Agents to Keep ChatGPT Fast

OpenAI says its internal system now investigates and fixes slowdowns with AI agents, and here's what that looks like.

This article was written by AI. It passed automated fact and quality checks, and our editor spot-checks published articles every day.

See AI World HQ first on GoogleAdd us as a preferred source and our stories show up higher in Google’s Top Stories.Add on Google

Have you ever typed a question into a chatbot and waited a few seconds for the first word to appear? That short pause is exactly the kind of problem engineers at AI companies work hard to prevent. OpenAI has published an inside look at how it now handles these slowdowns, and the approach shows where AI agents are heading inside large software teams.

What a "regression" is

A regression is when a change to software makes something worse than it was before, such as a page loading more slowly. Engineers call the normal, expected level of performance the baseline (the reference point you compare against). When performance falls below it, an alert (an automatic warning) usually fires.

OpenAI says its engineers used to dig through metrics, logs, and live machines by hand to find the cause. Over time the team built more than 30 performance tools to help, but the work still depended on people knowing where to look.

From manual digging to an automated workflow

OpenAI describes its progress in stages. First, engineers started asking Codex, its AI coding agent, to investigate slowdowns in code they didn't know well. Next, the team wrote down its investigation method as a reusable "skill" called perf-forensics, which tells Codex which tools to use and how to narrow down a problem.

Eventually the company moved these investigations to the cloud, so an alert could start one automatically. The catch was that the automation stopped at the diagnosis. A person still had to write the fix and check it after release. That gap is what the Performance Factory was built to close.

How the first workflow runs

The Performance Factory is a set of connected parts. According to OpenAI, it includes:

  • A task tracker where engineers define workflows using configuration files and written instructions for the agents.
  • A dispatch loop that hands each ready task to an agent.
  • Cloud execution that runs Codex in the cloud.
  • Subscriptions that wake an agent when something happens, such as a review comment or a finished deployment.

In the Regression Workflow, an alert opens a task. An agent investigates, records its findings in that task, and proposes a fix for an engineer to review. After the fix is released, another agent checks whether performance is back to its baseline. OpenAI says this first workflow now covers the whole process, from investigation to verifying the fix.

💬 Think of it like a relay race: each agent hands the baton to the next, and the task itself keeps the notes so nothing gets lost between handoffs.

Wrap-up

OpenAI says its Performance Factory turned a slow, manual process into a tracked, mostly automated one, handling 17 investigations a day on average in its first workflow. For you, the lesson is that AI agents work best when they follow clear instructions, keep good notes, and hand off to a person at the right moment. Your next step: the next time a tool feels slow, think about where in the process a clear record and a checkpoint would help most.

Keep reading

📬 The week’s AI, in your inbox

One friendly email every Sunday — the 5 stories that mattered, in plain English. No spam, unsubscribe anytime.

Was this helpful?

✦ Generated by AI in AI World HQ's automated newsroom, from official sources. Checked by automated fact and quality gates, and our editor spot-checks published articles every day. Spot a mistake? Use the buttons above.

← Back to all stories