A checklist for testing a voice agent before it takes real calls

Most problems in a new voice agent come from the same short list: the caller says something unexpected, switches language, stays silent, interrupts, or needs a person. You can find nearly all of them before launch.

Talk to it yourself

Call the agent in the browser and play difficult callers:

• Answer every question with one word.
• Interrupt it mid-sentence.
• Switch from English to Hindi halfway through.
• Go silent for ten seconds.
• Ask something completely off-topic.
• Ask to speak to a person.
• Give a wrong order number, then the right one.

Each surprise is a line to add to the prompt.

Run simulated conversations

Agent tests run whole conversations against the agent with a simulated caller who follows a persona and a goal ("a busy customer who wants to cancel", "an elderly patient who speaks only Hindi"), and check the result against criteria you set. Run them every time you change the prompt, so a fix for one case doesn't break another.

Check every tool and transfer

• Call each tool from the console's tool tester with valid and invalid arguments.
• Make a test call that triggers each tool and read the call timeline.
• Trigger a transfer and make sure the right person picks up with the summary.

Check the outputs

• Does each analytics field come out right on your test calls?
• Does the webhook reach your system, and does your system do the right thing with it?

Pre-launch checklist

1. Calling hours set.
2. The number attached, and the caller ID showing your business name.
3. The welcome message says who is calling.
4. Transfer targets set and tested.
5. Analytics fields and webhook tested end to end.
6. Enough credits for the first week.

Week one

Read or listen to a sample of calls every day for the first week. Look at transfers and calls that ended early: they show what the agent doesn't handle yet. Most agents improve the most in their first two weeks of real calls.