Which sufficiency failure modes should I test before deploying an agent?

Use this list to build or review the failure cases in an evaluation set for sufficiency decisions.

- **Presence-only checking:** the agent accepts a document because the type is right, ignoring age, entity or completeness. - **Policy version drift:** it applies the current requirement list to a case opened under an older one. - **Repeat requests:** it asks again for an item the customer already supplied. - **Vague requests:** the message says 'more documents needed' without naming the item, period or reason. - **Over-asking:** it demands items the requirement set marks as optional or substitutable. - **Silent conflict:** two documents disagree and the agent notices neither. - **Partial reply collapse:** after a half-complete reply it either proceeds or restarts from nothing. - **Missed escalation:** it continues on a case that policy says needs human review.

For each mode, ask for several cases where the correct behaviour is unambiguous and a few where it is arguable. Keep the arguable ones out of any pass or fail gate until your reviewers have agreed on the labels.