Home / Technology / OpenAI's AI Automates Code Fixes, Keeps ChatGPT Fast
OpenAI's AI Automates Code Fixes, Keeps ChatGPT Fast
19 Nov
Summary
- AI system automates investigation and mitigation of performance regressions.
- System tracked over 1,700 investigations in three months.
- AI agents are evolving into automated workflows, aiding engineers.

OpenAI has introduced the Performance Factory, an internal AI system designed to automate the investigation and mitigation of performance regressions in software. This system leverages Codex to streamline the process of identifying and fixing slowdowns, a critical task for maintaining services like ChatGPT.
The Performance Factory transforms manual debugging into automated workflows. Initially, engineers relied on manual checks, but the integration of Codex allowed for AI-assisted code analysis. The system evolved into a task-based architecture where AI agents systematically address performance alerts, from root cause identification to proposing fixes.
Key components include a task service, a dispatch loop for initiating AI sessions, an internal API for running Codex, and subscription services for external event integration. The Regression Workflow is a prime example, utilizing an "adversarial review" process to challenge initial diagnoses and ensure accuracy.
While the system has successfully identified root causes for hundreds of regressions and influenced numerous fixes, OpenAI aims to further automate remediation. Current efforts focus on refining agent-generated fixes and exploring proactive optimization strategies, such as identifying slowdowns missed by monitoring alerts.