Agent runtime

Voice Agents

AI agents you talk to — streaming speech-to-text, an agent turn, and streaming text-to-speech composed into a full-duplex loop where interruptions and latency are measured in milliseconds.

Operating principle

Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.

What it is

A voice agent is a realtime pipeline: microphone audio streams to speech-to-text, the agent processes the transcribed turn, and text-to-speech streams audio back — with barge-in (the caller interrupts playback and the agent stops) and end-of-turn detection making it feel like conversation. Two architectures compete: unified speech-to-speech models such as OpenAI Realtime, and composed pipelines that chain dedicated STT, an LLM agent, and dedicated TTS.

Why it matters

Voice remains the highest-volume channel in many operations, and its latency budget is unforgiving: the caller expects first audio in under a second. That budget forces architectural decisions — streaming everywhere, sentence-level TTS chunking, interrupt handling — that text chat never faces.

How we build it

A composed pipeline over WebSockets: streaming speech-to-text (Deepgram-class realtime STT) with tuned end-of-turn thresholds, the same governed agent runtime as text chat — same tools, same permissions, same traces — and sentence-chunked streaming text-to-speech with barge-in that cancels playback the moment the caller speaks. Keeping voice on the same runtime means the eval and governance story does not fork per channel.

Related resources