1. Voice agents

A voice agent is an LLM agent you talk to out loud, usually over a phone call, that listens, decides, calls tools, and answers in speech. Picture the phone line for a dental chain with 40 clinics, where patients call to book, move, or cancel cleanings. The agent behind that number runs the same loop you already know from Agent Architecture Patterns, with one difference: the caller is waiting on a live line. Speech comes in, the model decides, and speech goes out, all in under about a second, or the caller starts saying "Hello?"

Use a voice agent when people already call (appointments, order status, support) and the human queue is the bottleneck, or when hands and eyes are busy, as in a car. Skip it when the user has to review dense output, such as comparing ten insurance plans. A web page does that better, and a phone agent can text a link instead. Skip it too when the flow is a fixed menu of five options: a touch-tone menu is cheaper and more predictable than a model on the line.

This is an optional topic, asked in depth mainly at voice AI and contact-center companies, usually as "How do you build a voice agent that answers fast and handles interruptions?" A mid-level answer names the pipeline shape, a 1s target, filler speech (a short line spoken while a slow step runs), and confirming values before booking. A senior answer adds per-stage latency numbers, the endpointing trade-off (how long to wait before deciding the caller is done), what happens to conversation history when the caller interrupts, and when to mix pipeline types.