What changed

OpenAI described an internal workflow that starts from a performance alert. A Codex agent investigates and proposes a fix; an engineer reviews and deploys it. After deployment, an agent checks the affected metric against the earlier baseline.

Why it matters

The useful design is a closed loop between a real alert, human shipping authority and a measurable recovery check. A proposed patch does not by itself show that service performance recovered.

What to do

Replay a resolved latency incident with a known metric and baseline. Give a trial workflow read-only diagnostic access, keep deployment approval with an engineer, and compare its findings and recovery check with the actual incident record.

Caveat

This is an internal case study, not a product release or an independently replicated result. OpenAI discusses more autonomous shipping as future work, not the current authority boundary.