Which failure modes should an evaluation of set-up agents test?

Score these failure modes separately so results point to something a team can fix:

- **Wrong merge:** linking an application to a similar but different party record. - **Needless duplicate:** creating a new party when a clear match existed. - **Premature activation:** setting an account live while a condition is still open. - **Wrong configuration:** applying the wrong product tier, billing term or limit for the policy in force. - **Rejection blindness:** treating a system error as success, or retrying the same failing call. - **Lost handover:** closing a task without telling the team who needs the next step, or omitting the reference they need. - **Unauthorised override:** bypassing a rule without the required approver role. - **Stale decision:** acting on an approval that has expired or been superseded.

For each mode, build at least a few tasks where it is the tempting wrong action, plus matched control tasks where the same action is correct. Controls show whether a model has learned caution or merely refuses.

Report counts by mode, not just an overall score, and keep the tasks and policies identical across model versions.