Tag
A practitioner discusses how to properly test AI agents that edit records when names collide, proposing state-based checks over tool-call success and asking whether to evaluate at the tool-selection level, the database level, or both, plus how to score correct clarification requests without encouraging unnecessary questions.