OpenAI's Performance Factory automates ChatGPT performance regression detection and fixing
·2026.11.19 17:00
Key point
The system uses Codex to automatically investigate, diagnose, and propose fixes for performance issues, reducing manual engineering effort.
Details
- OpenAI's Performance Factory is an autonomous system using Codex to detect, investigate, and fix performance regressions in ChatGPT.
- The Regression Workflow has tracked over 1,700 investigations and identified 600+ root causes, averaging 17 automated investigations per day.
- 70% of root cause proposals from manual reviews were accurate, while 30% required further refinement or causal confirmation.
- The system evolved from manual analysis to a fully automated end-to-end agent workflow with four components: task service, dispatch loop, cloud-based Codex, and event subscriptions.
- An adversarial review process is included to prevent agents from stopping at the first plausible explanation, ensuring hypotheses are tested against contradictory evidence.
- The system supports self-repair by monitoring its own tool failures and generating fix tasks when errors impede multiple investigations.
- Future directions include proactive optimization by comparing performance data to find slowdowns before alerts trigger, and exploring agent autonomy for code review and merging.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.