Your agent should read a catalog and prepare a report. Halfway through, it requests an extra permission. Use the scenario to check whether the decision follows the original task.
A decision inside the session
On October 30, OpenAI announced auto_review in the Agents API beta for eligible projects. A model inspects supported requests, including some tool calls, and may approve or deny them. Review tokens are billed separately. Environment network restrictions and application-managed browser access requests follow their own rules.
What to check in your application
In an eligible test session, use a fictional catalog and a mock tool. Prepare an allowed read request, an ambiguous request and an out-of-scope change. Define expected application outcomes, then record review decisions, retries and billed usage. Also try stopping without automatic retry.
Keep business permission checks in the application before any real side effect. A review mismatch is a case where your integration should stop or ask for intervention.