Realtime Agents Integration

In a voice agent, the caller starts speaking while Sonic is mid-utterance; the stack keeps playing the synthesized audio over the caller.

On detected user speech (barge-in), immediately stop local playback AND cancel the in-flight TTS context so the agent yields the floor. Resume the dialog from the interrupted point. Treat interruption as a first-class control path, not an afterthought.

The agent pipes STT→LLM→Cartesia TTS but blocks on full LLM completion before starting synthesis, adding seconds of dead air.

Run the pipeline in a streaming/duplex fashion: begin streaming partial LLM text into /tts/websocket (chunked transcript on one context_id) so audio starts before the LLM finishes, while still capturing inbound audio for barge-in. Overlap stages to keep latency within the voice-agent budget.

End-to-end turn latency regresses; the team blames TTS without measuring per-stage contributions.

Instrument each stage (STT finalize, LLM first-token, TTS first-audio, network/playout) and attribute the turn budget per stage. Optimize the actual bottleneck. Treat any vendor latency figure as [REQUIRES-VERIFICATION] and rely on measured p95 stage timings.