Back to Tech Radar
AI Engineering & OperationsNov 19, 2026

Agent Incident Workflows Need Evidence for Every State

OpenAI’s internal Performance Factory case adds a useful incident-response lesson: define the evidence needed to move from a suspected cause to a reviewed fix and verified recovery.

Three connected abstract operational modules with evidence tiles and a smaller amber branch representing unresolved incident work.

The signal

An incident workflow becomes easier to audit when every change of status has an evidence requirement.

OpenAI's November 19 Performance Factory account offers a concrete example. Its internal system uses durable tasks and explicit transitions to coordinate Codex investigations, engineering review and checks after deployment.

The important question for builders is what permits a task to advance—and who owns it when the evidence remains incomplete.

What the case establishes

OpenAI describes a separate adversarial review that challenges a proposed root cause. Engineers review and ship fixes; an agent then compares performance with the baseline. Unchanged performance can return the task to investigation. Partial recovery can create a linked task for the remaining problem.

The company's manual review of 100 root-cause suggestions found 70% accurate. That is a vendor-reported diagnostic sample, not a general agent accuracy rate or a fix-success rate. The account describes an internal production system, not a generally available product. It also acknowledges that proposed fixes can introduce problems.

Why It Matters

Symphony made task orchestration and review boundaries visible. Our agent-evaluation article focused on evidence at meaningful milestones. Our GPT-6 Cyber analysis distinguished a finding from verified remediation.

This case gives us a reason to make incident status more precise. A convincing explanation, an accepted patch and a recovered service should each have their own acceptance criteria.

Builder takeaway

For an incident agent, define an evidence contract before increasing its authority. A practical starting point is:

State Proposed acceptance evidence
Cause proposed Observed symptom, timeline, supporting telemetry and competing explanations.
Cause reviewed A reviewer tries to falsify the explanation; unresolved counterexamples remain visible.
Fix ready A bounded change, tests, expected recovery signal and rollback owner.
Deployed The exact change and deployment are linked to the affected service.
Recovery verified A specified observation window and comparable baseline support the outcome.
Partially recovered Remaining impact has an owner and linked work; the original record states its limited closure scope.

These are Tech Radar's recommended checks, not a reproduction of OpenAI's implementation.

Choose the recovery window and acceptance threshold before the result arrives. Keep “diagnosis accepted” separate from “customer impact resolved.” If telemetry is missing or the checking tool fails, retain an explicit unresolved state.

We have not independently reproduced Performance Factory or measured its effect on other teams. The transferable lesson is to make each transition reviewable before treating a longer autonomous run as evidence of a better outcome.

Sources

How strong is this signal for builders?

Signal feedback is stored anonymously and used to improve Tech Radar editorial quality.

If this was useful, we can send you the next one.

We follow changes in software, AI, and infrastructure—and explain what they mean for people building and running systems.

Occasional Tech Radar emails from uniQubit. Unsubscribe anytime.

Sent by uniQubit · 1407 - 177 Nonquon Rd., Oshawa, ON L1G 3S2, Canada · info@uniqubit.ca
Privacy policy · Your confirmation email includes the controls to unsubscribe or change tracking.

Want more operational technology signals?

Follow uniQubit Tech Radar or contact uniQubit about a product, partnership, or operational software need.