How should we judge whether an agent set up an account correctly?

Two kinds of check work together for set-up tasks.

**Final-state checks** compare the records after the run with an expected state. Does exactly one customer exist for the applicant? Is the account linked to it, on the approved product code? Is the service activation state correct? Does a handover task exist for the right team? These checks are mechanical and repeatable, and they do not depend on how fluent the agent's summary is.

**Trajectory checks** look at how the agent got there, using the audit trail of tool calls. Did it search before creating? Did it read the relevant policy? Did it write in a valid order? Did it avoid actions outside its permission, such as editing a closed account without authority? An agent can land on the correct end state by guessing, or by making a destructive change and then repairing it. Trajectory scoring picks that up.

**Side-effect checks** confirm nothing else changed. Compare unrelated records before and after, and flag any difference.

**Honesty checks** compare what the agent says it did with what the audit trail shows. Claiming success after a rejected write is a distinct failure from not completing the task.

Report these separately rather than as one score. A single pass rate hides whether failures come from matching, policy, ordering, recovery or reporting, and those need different fixes in training or prompting.