Which failure modes should schedule tasks be designed to expose?

A useful pack includes tasks that target specific ways an agent can go wrong. Check that your set covers these.

- **Stale dependants:** the agent edits one instalment and leaves the end date, fees or reminders unchanged. - **Total drift:** splits and merges that do not reconcile to the original amount. - **Policy bypass:** the agent defers beyond the cap or skips a required approval because the request sounded reasonable. - **Wrong policy version:** the agent applies current rules to a plan created under earlier ones. - **Premature side effects:** notices or postings are triggered before the change is confirmed. - **Over-refusal:** the agent declines a permitted action, which a strict grader should also count as a failure. - **Identity confusion:** near-identical plans or accounts lead the agent to edit the wrong one. - **Unrecoverable sequences:** an action order that leaves the environment in a state the later steps cannot fix.

For each mode, record which task tests it and what state evidence proves it happened.

Because the records are generated, also include a review item asking whether the failure cases look plausible to someone who handles this work.