Playbook

Evaluate with a synthetic caller on the production path, personas including hostile callers, a rubric judge spot-checked by people, and latency measured at the client.

Barge-in must require real caller energy at the microphone over a recent window, not just a speech-start event; the agent's own voice and line noise both trigger false interruptions.

Before any consequential action, the guard must find a recent, plain, action-specific yes in the caller's latest transcribed utterance, never trusting the model's claim that consent was given.

Rules that must always hold belong in tool code and gate checks, not in prompt text; the prompt decides what to say next, the code decides what may happen.

Each extra tool the model chains inside one turn is another language-model round trip of roughly a second; return the data the next step always needs inside the previous tool's result.

A hosted voice agent can go silent for a minute or more mid-call; a well-armed watchdog detects it in seconds and recovers exactly once, without repeating any action the caller already triggered.

Voice-to-voice latency is end-of-turn detection plus one or two model calls plus time to first audio; under about 800 ms feels conversational, over about 1.2 s feels like an IVR.

Unit tests and scripted harness runs cannot hear audio; a voice agent is verified only after a real browser or phone run, golden recordings, and one live call from a network like the audience's.

An injected filler is a second speech stream; it collides with the caller, confuses reply completion, and if it counts as a reply it hides the stall it was meant to cover.

A fixed silence timeout cannot tell a pause from an ending; it splits turns on mid-sentence pauses, holds turns open on trailing digits and noise, and adds a fixed delay to every reply.

A text-and-timing harness cannot observe playback stutter, overlap or network jitter; the only evidence of a working voice agent is a real client on a realistic network with counters.

A rule that lives only in the prompt is a preference the model usually honours; a control is code that refuses regardless of what the model decides.

The advertised per-minute price is usually the orchestration fee alone; speech, language model, voice, telephony, numbers and tooling roughly double to quintuple it.

Anything placed in a tool result is candidate speech; internal notes must be named as not-for-speech, kept terse, and backed by a persona rule, or the agent will read them to the caller.

A demo the vendor drives is a rehearsed recording of the agent at its best; only a demo the buyer drives, with an adversarial persona, measures the agent.

Every speech-recognition system tested fabricates words on speech-free audio at some rate; a word whose timestamps fall outside the audio is fabricated by construction.