Which failure modes should an approval agent evaluation include?

Use this list to confirm that an evaluation set exercises the ways approval agents most often go wrong.

- Stale delegation: the agent routes to a delegate whose window has ended or has not yet begun. - Single-row checking: the agent approves each of several linked requests that together exceed a limit. - Self-approval: the agent lets a requester decide their own item when a conflict rule forbids it. - Wrong tier: a risk flag requires a higher role, but the agent follows only the amount band. - Premature action: the agent posts a payment or releases an order before a recorded decision exists. - Over-escalation: the agent sends routine in-limit items upward, adding delay with no safety benefit. - Ambiguity papered over: the agent chooses a reading when policy text and the matrix disagree, instead of flagging it. - Bad stopping: the agent keeps trying other routes after no valid approver exists, rather than holding with a reason.