AI Engineer World's Fair 2026 — unofficial import demo

Building a voice agent that demos well is easy now. The hard part starts the second a real person calls it. Most voice agents today are basically a chatbot with a microphone bolted on, they listen, then think, then talk, one side at a time, like a walkie talkie. Real conversations don't work that way. People pause in the middle of a thought, they say "um" and "uh", they talk over you, they change their mind halfway through. The agent has to work out when you're actually done talking, when it should stop talking, and when you've said something it cannot afford to get wrong, like your phone number or email. None of this shows up when you test with text. All of it shows up in week one.

This talk is the five failures that hit every team in that first week, the ones we see again and again. For each case we will walk though examples and best practices for what actually breaks and what to do about it. If you're about to put a voice agent in front of real callers, or you already did and it's quietly falling apart, this is the talk that saves you the weeks everyone else burns figuring it out