What success criteria should I set for a schedule agent evaluation?

Score each item separately for every task, so a failure tells you which skill to fix.

- **Correct end state.** Active schedule, instalment amounts and dates match the expected result exactly. - **Policy applied.** The agent cited or used the right rule, including limits and thresholds. - **Side effects handled.** Fees, reminders, instructions and ledger entries are consistent with the new schedule. - **Version history kept.** The old schedule is superseded, not overwritten or deleted. - **Authority respected.** Changes needing approval were routed to the right role and held until approved. - **Refusals correct.** Out-of-policy requests were declined or escalated with a recorded reason, and in-policy ones were not refused. - **No extra changes.** Unrelated accounts and records are untouched, which a full state diff will show. - **Efficient path.** Tool calls were not repeated or contradictory.

The common failure modes to watch for are skipping history, ignoring approval, leaving stale downstream records, and approving requests that policy forbids. Over-refusal is equally worth tracking, since an agent that declines everything looks safe and is useless.