OpenAI Performance Factory: Codex Agents Ran 1,700 Investigations to Keep ChatGPT Fast

OpenAI described its Performance Factory, an internal system where Codex agents investigate ChatGPT performance regressions, propose fixes and verify recovery, with engineers approving every change. Its first workflow ran more than 1,700 investigations and root-caused over 600 cases in just over three months.

Christopher Ort
By Christopher Ort · 3 min read
Share
An engineer seen from behind in a dim operations room with monitors showing blurred graphs.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. How it works
  6. Why it matters
  7. What's still open
  8. Sources
  9. i10x Perspective

OpenAI has described the Performance Factory, an internal system in which Codex agents investigate slowdowns in ChatGPT, find the cause, propose a fix and later check whether the fix worked. In just over three months, its first workflow tracked more than 1,700 investigations and found the root cause in over 600 cases, OpenAI said on 19 November 2026.

i10X Quick Take

Quick Take

  1. What it is: a cloud system where Codex agents run performance engineering workflows end to end, with engineers approving what ships.
  2. Scale: more than 1,700 investigations and over 600 root causes in just over three months, about 17 automated investigations a day.
  3. Accuracy: in a manual review of 100 root-cause suggestions, OpenAI graded 70% as accurate.
  4. Checks: a separate “adversarial” Codex session tries to disprove each diagnosis and each proposed fix.
  5. Humans stay in: an engineer reviews every pull request and decides whether to merge it.

In plain English

When ChatGPT gets slower, someone has to figure out why, often by digging through charts, logs and recent code changes. OpenAI now lets AI coding agents do most of that detective work. One agent finds the likely cause, a second agent tries to poke holes in it, and then an agent writes a fix. A human engineer still decides whether the fix goes live, and an agent checks afterwards that speed is back to normal.

What happened

In a post by Okay Zed and Cyrus Forbes, members of OpenAI’s technical staff, the company wrote that “the staggering pace of agentic development has made traditional approaches to performance engineering unsustainable.” Its answer is the Performance Factory, “an autonomous system that uses the full capabilities of Codex and our models to find and fix regressions.”

The first workflow, called the Regression Workflow, opens a task when a performance alert fires, for example a drop in ChatGPT latency or availability. The task moves through the stages needs investigation, root caused, fix proposed and resolved.

VentureBeat, which received the engineering report as an exclusive, reported that of the more than 600 cases with a root cause, around 400 reached a proposed fix, about 50 shipped fixes were influenced by or matched agent suggestions, and around 15 shipped fixes were exact pull requests written by Codex.

Background

OpenAI says its use of Codex went through six stages, from engineers working by hand with more than 30 internal performance tools, to Codex as a partner, to a shared skill called perf-forensics that captures the team’s investigation playbook, and finally to the factory. It fits a broader push at the company: OpenAI recently reported that AI agents now finish 46% of multi-hour research tasks on their own.

How it works

  • Task service: engineers define workflows with configuration files and Markdown prompts that list states, allowed transitions and which steps need a human.
  • Dispatch loop: it checks which tasks are ready and starts a Codex turn with the instructions for the current state.
  • Codex in the cloud: agents run inside OpenAI’s network under a service identity, with access to code, metrics, logs, traces, CPU profiles and memory snapshots.
  • Subscriptions: waiting tasks wake up on events such as a GitHub review comment, a failed CI check or a finished deployment.
  • Adversarial review: early agents “sometimes stopped at the first plausible explanation,” OpenAI wrote, so a second session with fresh context now rechecks assumptions and tests counterfactuals.
  • Self-repair: an agent scans recent runs every day for broken tools or missing permissions and opens tasks to fix them.

Why it matters

For engineering teams, this is one of the most detailed public accounts of agents doing production operations work, including where they fail. OpenAI says “getting fixes right has been harder than finding the cause,” with some proposed changes more complicated than needed or introducing new issues.

For developers using Codex, the patterns are reusable: split big jobs into focused agent turns, keep task state outside any one agent’s context, and test skill or model updates against past incidents before rolling them out.

For ChatGPT users, OpenAI says that when the service slows down, “there’s a good chance the Performance Factory is already figuring out why.”

What’s still open

  • OpenAI is “exploring how agents could use our other agentic review tools to identify low-risk changes and eventually ship those fixes without a human in the loop.” When and with what safeguards is not stated.
  • The 70% accuracy figure comes from OpenAI’s own grading of 100 cases.
  • Whether OpenAI will offer the factory pattern as a product for Codex customers.

Sources

  • OpenAI, “How OpenAI’s Performance Factory keeps ChatGPT fast,” Okay Zed and Cyrus Forbes, 19 November 2026
  • VentureBeat, exclusive report on the Performance Factory, 19 November 2026

i10x Perspective

The interesting part is not that agents can debug, but how OpenAI fences them in: explicit states, a second agent paid to disagree, and a human at the merge button. That is a template many teams will copy. More on agentic development in Coding.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    GPT-6.1 Terra

    1.05M context · $1 / $5 per 1M tokens

    openai.com (source, opens in a new tab)
  2. Release

    GPT-6.1 Sol

    1.05M context · $2 / $10 per 1M tokens

    developers.openai.com (source, opens in a new tab)
  3. Release

    GPT-6 Sol & Luna

    1.05M context · $2 / $10 per 1M tokens

    Two GPT-6 reasoning models. Price shown is GPT-6 Sol (≤272K prompt); GPT-6 Luna costs $0.10 / $0.50.

    developers.openai.com (source, opens in a new tab)
  4. Release

    GPT-6 Astra

    1.05M context · $10 / $50 per 1M tokens

    OpenAI's most capable model for reasoning, coding, computer use and research.

    openai.com (source, opens in a new tab)
  5. Release

    GPT-5.6

    1.05M context · $4 / $20 per 1M tokens

    GPT-5.6 family: Sol (frontier), Terra (balanced) and Luna (high-volume).

    developers.openai.com (source, opens in a new tab)
Full ChatGPT timeline ›

Editor's picks

  1. What Comes After AGI? Google DeepMind Maps Four Paths to Superintelligence
  2. AMD and Unsloth Bring NPU Support to Unsloth Desktop, Letting 100+ AI Agent Apps Run Locally on Ryzen AI
  3. Perplexity Brain Now Writes Its Own Agent Skills, With 80,000 Created for 20,000 Users
  4. OpenAI Performance Factory: Codex Agents Ran 1,700 Investigations to Keep ChatGPT Fast
  5. OpenAI GPT-Live-1.1: New Voice Model Improves Transcription and Accents, Adds Cheaper Standby