OpenAI's software factory closes the loop from CI to production

The Pragmatic Engineer reports that OpenAI has connected Codex to an end-to-end software delivery loop: agents gather internal context, implement and test changes, run specialist reviews, classify risk, monitor rollouts, and turn production regressions into proposed fixes. The notable shift is not code generation alone, but a harness that carries a change through CI and into production feedback. Why it matters: When agents increase change volume, review, CI capacity, deployment safety, and observability become the limiting system. OpenAI reportedly saw roughly 10x load growth in parts of its build-test-deploy infrastructure, making risk-based gates and per-change monitoring more important than raw generation speed.

Try this: For one low-risk change, require an agent to record the diff, CI result, risk label, and deployment trace. Review that evidence before rollout, then compare the predicted success signals with what production actually showed.