The hardest number in a voice agent is 800 milliseconds.

That's roughly how long a pause can last in a phone conversation before the person on the other end thinks something is wrong. They say "hello?" or they hang up.

Inside that window you have to transcribe what they said, decide what it means, work out the reply and turn it back into speech. Four steps, each with its own latency, plus whatever the phone network adds on top.

Text agents get to think for five seconds and nobody notices. On a call, thinking is silence, and silence reads as a broken line.