Working Surfacelink · 21 Nov 2026
01
A railway line with three signal boxes in a row, where a mechanical arm at the second box drops a red flag to halt a small goods wagon until a signalman seen from behind at the last box pulls the lever that lets it through, the whole scene drawn with wide empty margins on every side.
OpenAI

How OpenAI’s Performance Factory keeps ChatGPT fast

Okay Zed and Cyrus Forbes · 19 November 2026

Company story · critique & review · agents · Human approval

Read the original
Takeaway

OpenAI engineers built a system in which Codex AI agents investigate and fix ChatGPT slowdowns, a second agent challenges each diagnosis, and an engineer decides what is merged.

Summary

Okay Zed and Cyrus Forbes, engineers at OpenAI, describe the Performance Factory, a system that keeps ChatGPT fast. AI agents help ship code so quickly that engineers could no longer investigate every slowdown by hand. Codex, OpenAI's coding AI agent, now investigates each performance alert, proposes a fix and checks recovery, while an engineer reviews each fix before it is merged.

Key points
  • In just over three months, the system tracked more than 1,700 investigations and found the root cause of over 600 slowdowns.
  • A separate Codex session tries to disprove each diagnosis; in one case it reopened an investigation that had named a cause which did not explain the continuing failures.
  • In a manual review of 100 diagnoses, the team graded 70 percent as accurate and 30 percent as needing more work.
  • The team tests changes to the agent's instructions by replaying past production incidents, with the incident write-up hidden from the agent.
Implication

An engineering team can let AI agents carry routine investigation when each step has a separate check and a person decides what is merged. That is an inference and is not the authors' claim. The article is OpenAI's account of its own product, Codex.

Plays
Engineering

Test changes to an AI agent's investigation instructions by replaying past production incidents, with the answer hidden from the agent.

Engineering

Have a second AI agent session try to disprove an AI agent's diagnosis before any fix is written.

Derived by Working Surface from the article; more in the Playbook. Source line: We want engineers to spend their attention on decisions rather than reconstructing the agent’s work.

Source issue

21 November 2026: as AI agents take on investigation, design and drafting, the human work moves to choosing what the agents read and checking what they produce before it reaches customers.