Test it like a stranger

You are the worst possible tester of your own agent, because you already know what it is for and you unconsciously ask the questions it can answer.

So do this instead. Write down five questions before you open the chat: three a real customer would ask, one a customer would ask that you know the agent cannot answer, and one that is completely off-topic. Then ask all five, in order, without helping.

The two that matter are the last two. The fourth tells you whether your escape hatch works. The fifth tells you whether your boundary holds. If the agent answered either one by making something up, the instruction is not finished — and you have just learned exactly which of the three parts is missing.