Configure a voice agent

Configure the agent as a complete system. A strong prompt cannot compensate for an incompatible voice pipeline, and low latency cannot compensate for missing business rules.

Start with the call design

Before changing controls, write down:

• the caller's likely intent;
• the one primary outcome the agent should achieve;
• facts it may state and facts it must not invent;
• information it must confirm before taking action;
• when it should transfer or end the call; and
• what success and failure look like in a call log.

This becomes the basis for the system prompt and test cases.

Prompt Editor

Use the system prompt for identity, primary outcome, conversation rules, source hierarchy, action policy, escalation, and boundaries. Keep spoken responses concise with explicit instructions such as “ask one question at a time.”

Use exact {{variable_name}} markers for configured variables and [[Function: action_name]] markers for attached actions. The editor warns about missing variables, unattached functions, and disconnected integrations, but warnings are not a complete safety review.

See Write an effective system prompt.

Choose the pipeline architecture

Modular

Modular runs speech recognition, a text model, and speech synthesis as separate stages. It offers broader component choice and more granular speech controls.

Realtime

Realtime uses a provider-native audio conversation path. It can produce a more direct conversational exchange, while available voices and controls depend on the selected provider.

Wirevox preserves separate Modular and Realtime selections in the draft. Switching architecture can reactivate previously stored settings, so review the complete configuration after every switch.

See Compare Modular and Realtime and Review configuration compatibility.

Configure the editor sections

Section	Configure	Review carefully
General	Timezone and compatible conversational model.	Time-sensitive behavior, model availability, and cost.
Variables	Stable business values referenced by the prompt.	Missing values and accidental secrets.
Call Settings	Inbound answer delay.	Carrier experience and expected pickup timing.
Language & Voice	Primary language, voice, optional Modular secondary language, and supported speech recognition.	Provider and language compatibility.
Voice Delivery	Personality, expressiveness, reactions, pitch, pronunciation, and provider-specific options.	Sensitive-call tone and unsupported controls.
Audio Environment	Optional ambient sound and presence.	Intelligibility and distraction.
Knowledge Base	One agent-level workspace knowledge base.	Source freshness and grounding rules.
Conversation Settings	Response timing, speed, interruption, filler words, silence, hangup, and duration.	Realistic noisy-call behavior.

Controls are capability-aware and can be hidden or restricted by architecture, model, provider, language, or voice.

Keep policy separate from mechanics

Use the prompt to define business policy: when to ask, confirm, act, transfer, or refuse. Use configuration controls for audio mechanics: recognition, speaking pace, silence, interruption, and duration.

For detailed tuning, see Voices, languages, and speech and Conversation behavior and call controls.

Save and cost behavior

The editor autosaves into the draft after a short delay and shows its save state. Opening a test, publishing, navigating away, or changing status attempts to flush pending changes first.

The header estimates the draft's voice cost in credits per minute and can break it down by platform, speech recognition, model, and speech synthesis. The published configuration can have a different rate until the draft is published.

Configuration review checklist

• Prompt names the correct organization and purpose.
• Timezone, language, and pronunciation match callers.
• Architecture, model, recognition provider, and voice are compatible.
• Greeting does not duplicate the prompt or promise an unavailable action.
• Silence, interruption, and call limits match the use case.
• Cost estimate is acceptable.
• Save state confirms the latest edit succeeded.