Describe the business and the generator returns a starter scenario set split three ways: calls the agent must handle end to end, calls it must escalate, and calls it must refuse. That split is the whole structure of an eval suite, and it is the step almost every pilot skips.
The caller who changes their mind after a slot is offered. The caller who gives an address outside the service area halfway through. The caller who is calling about a job that already happened. All three are common, all three break naive agents, and none appears in a vendor demo.
Run each scenario before every prompt or knowledge change and record pass or fail with the transcript. A failure blocks the change. That gate is the difference between an agent that improves over time and one that oscillates as each fix quietly breaks something else.