ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
OpenAI's Performance Factory Goes Live: Codex Agents Fix Performance Regressions on Their Own, Handling 1,700 Cases in Three Months

OpenAI's Performance Factory Goes Live: Codex Agents Fix Performance Regressions on Their Own, Handling 1,700 Cases in Three Months

AI information • Admin • • 21 views

OpenAI's Performance Factory turns fixing performance problems into an agent assembly line: on November 19, 2026, OpenAI said in its official blog that the Codex-powered automation system has been running internally for more than three months, handling over 1,700 performance regression investigations and finding the root cause in more than 600 of them. Engineers no longer chase every alert — they step in only at the key decision points.

How it goes from alert to verified fix

The first workflow to go live is called the Regression Workflow: when a performance alert fires, a task is created automatically, and an agent first pins down when the metric got worse and which requests were affected, then checks deployments and configuration changes and reads the relevant code to find the cause. Once the diagnosis passes, the agent writes a fix and opens a pull request, an engineer reviews and merges it, and after deployment another agent checks whether performance has returned to baseline before the case is closed. Four parts hold the chain together: a task service with configurable workflows, a dispatch loop that assigns work, an API that runs Codex in the cloud, and subscriptions that wake agents when events such as code reviews or completed deployments happen. Engineers distilled their debugging experience into a skill called perf-forensics, and agents repeat the same playbook, handling 17 investigations a day on average. OpenAI Explains the Codex Sandbox: No Containers, Native System Mechanisms Keep the Agent Caged describes how Codex runs inside a restricted environment — the precondition that lets this kind of automation run with confidence.

Seven in ten diagnoses are right; landing the fix is the hardest step

OpenAI does not hide the system's weak spots. In a manual review of 100 root-cause suggestions, 70 were graded accurate; the other 30 needed more work or had not fully established causality. The team added an adversarial review to catch this: a separate Codex session tries to disprove the diagnosis, checking, for example, whether similar requests that were not exposed stayed healthy, or whether the metric recovered after a rollback. Fixing is harder than finding the cause: agent-proposed changes are sometimes more complicated than necessary, or introduce new issues, and trade-offs such as adding a cache still need an engineer's call. Every proposed fix goes through the same adversarial review, and only low-risk changes may one day ship without a human in the loop.

Next, OpenAI is building an evaluation loop: comparing the fixes agents propose with the changes engineers ultimately merge, building evaluation sets, improving the workflows, and using synthetic performance tests so agents can test a fix before submitting it. If this approach holds up, performance engineering shifts from 'people hunting for problems' to 'agents hunting for problems, people making the calls' — for other teams running services at scale, that working pattern may matter more than any single number.

Recommended Tools

More