OpenAI described its Performance Factory on November 19, 2026. This is the date of the public engineering disclosure, not the system's initial deployment.
The disclosure describes a persistent investigation workflow
The post connects a task service, dispatch loop, cloud Codex execution and event subscriptions. Findings persist beyond an individual agent context. A separate review session challenges diagnoses, engineers decide which changes ship, and the workflow checks production recovery after deployment.
OpenAI reports more than 1,700 investigations and over 600 root-caused cases in just over three months. Those are reported operational counts, not independent productivity or correctness measurements. The authors acknowledge that proposed fixes can be overly complicated or introduce issues.
Keep evidence of recovery separate from evidence of a patch
Engineering interpretation: a proposed change, a deployed change and a recovered measurement support different conclusions. Keeping these states explicit would prevent a workflow from using the existence of a patch as its only completion condition.
A review record should also preserve which observations were available when the diagnosis was made. Later evidence can justify a new conclusion without being silently treated as information the original investigator already possessed.
A proposed exercise with a deployed but ineffective fix
Original exercise, not executed: generate a synthetic latency trace containing a known delay. Define four states: investigated, fix proposed, deployed and recovery checked. Supply a second trace in which the proposed patch has deployed but the delay remains.
The expected outcome is that the task stays unresolved and returns to investigation. A separate fixture removes the delay, allowing closure only after its measurements pass the declared check. Preserve timestamps and prevent the initial diagnosis step from reading later recovery fixtures. This is a small state-transition exercise, not a reproduction of OpenAI's service or its reported results.
Sources accessed November 21, 2026. Views are my own and do not represent any employer, past or present.
Sources and evidence
Tie task closure to observed recovery, while preserving separate diagnosis, review and deployment records.