OpenAI details Performance Factory: agents investigate, engineers decide what ships

Read the original article →

OpenAI’s November 19, 2026 engineering post describes Performance Factory, an internal Codex-powered workflow for investigating slowdowns. It is an account of OpenAI’s own infrastructure, not an announcement that customers can enable the same service.

From an alert to a verified recovery

According to OpenAI, an alert opens a task; agents investigate and propose changes; an engineer reviews and decides what ships. After deployment, an agent checks whether the affected performance returns to baseline. The task retains findings across those handoffs.

OpenAI says a separate agent challenges a proposed diagnosis rather than accepting the first plausible explanation. In its manual review of 100 root-cause suggestions, it graded 70% accurate; the remainder needed refinement or did not establish causality. This is a vendor-reported sample, not an independent benchmark or a guarantee for other codebases.

MintedBrain analysis: borrow the evidence loop

For a smaller team, the useful starting point is a narrow investigation with a measurable finish line. Record which request became slower, when it changed, and what an unaffected comparison looks like. Ask the agent for evidence that could disprove its explanation before asking it to patch the code.

Illustrative example: a checkout endpoint slows after a release. Compare affected and unaffected requests, inspect the relevant change, and measure the same endpoint after a reviewed fix. A merged patch alone does not establish recovery. Keep the result and remaining uncertainty in the issue so the next investigation can build on it.

This workflow advice is MintedBrain analysis. We have not tested Performance Factory or reproduced OpenAI’s results.

For the access boundary around agent commands, read our Codex sandbox summary. Explore tutorials for related learning workflows.

Primary source: OpenAI’s engineering post, November 19, 2026. Coverage checked November 22, 2026.

References

This article was originally published at OpenAI. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News