Trust is built in the first two seconds

People decide whether they are talking to something competent almost immediately. A half-second of dead air, a greeting that runs too long, a voice that talks over them — any one of these tells the caller they are dealing with a machine that is not really listening, and the conversation is effectively over.

So most of the engineering work is not in the language model at all. It is in the plumbing: how fast audio moves, how quickly the agent detects that the caller has started speaking, and how gracefully it yields when they do. Shaving latency is unglamorous work with an outsized effect on how human the agent feels.

Knowing when to stay quiet

The hardest behaviour to get right is restraint. Someone describing a car accident will pause to collect themselves, and an agent that treats every pause as its turn to speak becomes unbearable. Good agents wait. They let silence sit slightly longer than feels comfortable, because the person on the other end is often not finished.