OpenAI reveals Performance Factory: Codex agents that investigate ChatGPT slowdowns

OpenAI revealed on Wednesday (November 19) the Performance Factory, an internal system that uses Codex agents to automatically investigate and fix slowdowns in ChatGPT — performance engineering work that used to be manual. The reveal came in an exclusive engineering report shared with VentureBeat and a thread from the @OpenAIDevs account on X. In just over three months, the system has already run more than 1,700 investigations, identified the root cause in more than 600 cases and proposed around 400 fixes.

From alert to verified fix

When monitoring detects a performance regression — a service that starts responding more slowly after a change — the system opens a task and a Codex agent investigates: when the slowdown began, which requests were affected, which code changes and feature-flag changes happened in the period, cross-checked against logs and traces. The agent follows a fixed method, the reusable 'perf-forensics' skill, which reproduces the performance team's procedure. Then a second Codex session, unaware of the earlier work, tries to refute the diagnosis: about 25% of diagnoses are challenged in this adversarial review. If the diagnosis survives, Codex writes the fix and opens a pull request — and a human engineer decides whether it goes in. After deployment, another agent checks whether performance is back to normal; if recovery is partial, a new task is opened.

What the agents have delivered — and what they have not

The numbers show a narrow funnel: of the more than 1,700 investigations, just over 600 reached a root cause and about 400 produced fix proposals. Of those, roughly 50 fixes that reached production were influenced by the agents, and only 15 were pull requests written exactly by Codex and accepted as written. In other words, the system accelerates human work rather than replacing it: the final decision stays with engineers. OpenAI itself says it is still short of what it wants to achieve, and names as next steps comparing versions to find improvements and studying whether agents may one day apply simple, low-risk fixes without human approval.

The system has four components — a task tracker, a dispatch loop, cloud execution via the Codex API and event triggers — and the agents operate under a dedicated service identity with restricted permissions. It is a picture of how coding agents are moving from isolated tasks to orchestrating entire engineering workflows, with human review at the end of the line.

Image: OpenAI symbol ('Blossom', in use since February 2025). Credit: OpenAI, via Wikimedia Commons.

Sources

Transparency note: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and internet sources, and reviewed before publication. Please consult the original sources to confirm the information.

By Muse