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

Cover these failure modes with at least one task each, and tag results by mode rather than only pass or fail.

- **Duplicate creation.** The agent creates a new record when a near-match exists, or merges when policy says to stop. - **Wrong product or terms.** The agent selects a similar but incorrect product code, plan or billing cycle. - **Out-of-order writes.** A dependent record is created before its parent, leaving orphans or invalid links. - **Silent partial completion.** A later write fails, and the agent reports success or leaves no marker of the incomplete state. - **Skipped policy branch.** A required second check, referral or approval is bypassed. - **Missing handover.** Records are correct but the task or note for the next team is absent or addressed to the wrong team. - **Invented data.** The agent fills a missing mandatory field with a plausible value instead of asking or pausing. - **Overreach.** The agent changes records outside the approval, or uses a tool beyond its permission.

For each mode, decide in advance whether the correct behaviour is to complete, ask, escalate or refuse, and record that in the answer key. Add a few clean cases as well, so you can see whether training for caution makes the agent refuse work it should have finished.