What failure modes should I test for in agents doing booking tasks?

Test for the failures that cost money or break handovers, not only for wrong field values.

Common failure modes:

- **Premature activation:** the agent sets the facility to ACTIVE while a condition is open. - **Silent correction:** the agent overwrites a differing rate instead of querying credit. - **Stale approval:** the agent books against an approval that has since been amended. - **Missing link:** security or a co-borrower is not attached to the facility. - **Schedule drift:** the repayment schedule is built before terms are final and is not rebuilt. - **Over-escalation:** the agent raises a query on every case, which is safe but unhelpful. - **Invented values:** the agent fills a missing field with a plausible figure.

Each mode needs a case where the correct behaviour is clear and a grader can detect the wrong one from the final state or the action log. Include matching clean cases, where the right move is simply to proceed, so you can measure over-escalation as well as under-escalation.

When you describe the release to customers, say which modes the cases cover. Ask in the pilot for a mapping from each case to the failure mode it is meant to expose.