VOICE AGENT DESIGN PATTERNS: SCRIPTS, GROUNDING, BARGE-IN, AND HONEST WARM HANDOFF

A production voice agent resolves the calls that follow scripts, grounds every factual claim in your systems of record, handles barge-in without talking over the caller, and hands complex calls to humans with transcript and context attached. Anything less becomes a frustration amplifier with a nice voice.

Pattern 1 - Scripted flows with escape hatches

Status checks, bookings, and verifications belong in explicit state machines. Escape hatches ('speak to a person', low confidence, repeated failure) must be first-class transitions, not afterthoughts. See /products/voice.

Pattern 2 - Grounded dialogue, not improvisation

Answers come from Knowledge or live APIs - never from model prior alone. When retrieval fails, the agent says so and offers transfer. That single honesty rule prevents the worst class of voice incidents.

Pattern 3 - Barge-in and latency budgets

Callers interrupt. Design for partial ASR, cancel in-flight TTS, and keep tool-call latency inside a budget the ear forgives. Measure time-to-first-audio and time-to-handoff separately from 'model tokens per second.'

Pattern 4 - Warm handoff as a product feature

When a human takes over, they get transcript, intent, entities already collected, and suggested next steps. Cold transfers teach callers to bypass the agent forever.

Pattern 5 - Total QA, not sample QA

Score every call against your rubric automatically; sample only for calibration. The metric that matters is resolved-without-frustration - not containment rate alone.