A voice agent that handles the overnight queue

What we built

The agent sits on the operator's existing SIP trunk. Nothing about the published numbers changed — no porting, no new carrier, no change to daytime routing. Out-of-hours calls that previously hit voicemail are answered instead, and the daytime path is untouched.

1. Answer and transcribe. Streaming speech-to-text, so the agent has partial text while the caller is still speaking rather than after they stop.

2. Ground the turn. The agent calls the operator's own tracking and booking APIs through typed tools. It has no separate copy of the data and no cache, so it cannot be stale in a way the operator's own systems are not.

3. Decide, then speak. A turn either completes, asks one clarifying question, or hands over. There is no fourth branch, which is what keeps the failure mode legible.

4. Hand over with the context attached. An escalation writes the transcript, the caller's identity and the tool calls already made into the morning queue.

Barge-in is a feature, not an error

Callers interrupt. On a phone line that is normal conversational behaviour, and an agent that treats it as an error condition talks over people — which is worse than the voicemail it replaced.

So interruption handling was specified up front rather than tuned afterwards: speech from the caller stops synthesis immediately, the partial utterance is kept, and the agent resumes from what it had rather than restarting its turn. The evaluation set scores it directly.

The latency budget

Turn-taking latency is the number that decides whether a voice agent feels like a conversation, and it is a budget spent across the whole hop — not a property of the model. It was agreed during Discover, annotated at every stage, and measured end to end in the evaluation harness rather than at the model call, which is the measurement that flatters.