Last week I built an AI caller for a client. Its job is to ring their leads and set up a quick call with their account manager. It took four test calls before it sounded like a person. Not one of the fixes was a smarter model.
Vapi is a platform for building voice agents. You give it a prompt plus a voice and it handles the phone call. Claude Code changed the agent through Vapi's API after every test call and saved each version as JSON, so I could roll back.
If you've built a voice agent, what did it do on its first test call that made you wince?