You already know: agent loops and tool calling (D9.1, D9.3); speech recognition and synthesis basics (D18.3); time to first token and streaming (D12).
Why this matters for Atlas. In text, a 3-second answer feels fine. In conversation, a pause longer than about a second feels broken, and people start talking again, which the agent then has to handle. A voice agent is an agent whose every design choice is shaped by latency and turn-taking: stream every stage, decide quickly but correctly when the caller has finished, stop speaking the moment they interrupt, and confirm before doing anything irreversible because the caller can't see a screen. This lesson builds that, piece by piece.
1. The loop and the latency budget
The question. How fast does a voice agent have to respond, and where does the time go?
Start from nothing. In human conversation, the gap between turns is typically a few hundred milliseconds. A practical target for a voice agent is under about 800 ms from the caller stopping to the agent starting to speak. The classic cascade has four stages: