Agents overview

An agent combines a conversation engine, voice, instructions, knowledge, call controls, and post-call analysis. Provider-specific settings remain isolated, so switching an engine does not silently copy incompatible values.

Choose an engine

Engine	Pipeline	Best for	Key controls
Classic Pipeline	Speech-to-text → LLM → text-to-speech	Broad provider and voice choice	TTS provider, voice, language, speed, transcription languages
Vani Ultra	Whisper → Gemma → OmniVoice	VaniAgent-managed multilingual voice calls	Vani Ultra voice, vocabulary, language, generation steps
Ultra Realtime	Gemini Live speech-to-speech	Low-latency natural conversation	Gemini voice, realtime language, provider-specific prompt

Configuration map

1. Identity — name, active state, assigned phone number, and optional avatar.
2. Conversation — fixed or variable greeting, fallback greeting, system prompt, and engine-specific prompt.
3. Knowledge — attach one or more published knowledge bases.
4. Speech — voice, language, speed, transcription languages, and pronunciation vocabulary.
5. Call behaviour — interruptions, background music, and hard call-duration limit.
6. Analysis — default, structured, or advanced post-call extraction.
7. Advanced flow — optional flow PDL with off, shadow, or active runtime mode.

Caution

Always test after changing the engine, greeting, prompt, knowledge, or assigned number. A saved configuration can still be unsuitable for real callers.

Agent lifecycle

Create the agent, configure its engine, test it in the browser or with a controlled call, attach a phone number, then activate it for inbound calls or campaigns. Deactivate an agent before major production changes if it is already receiving traffic.

Post-call analysis

• Default produces the standard call summary and outcome.
• Structured extracts explicitly configured fields for predictable workflows.
• Advanced uses a custom analysis prompt and configuration.

Analysis is separate from the live conversation prompt. Changing analysis instructions does not alter what the caller hears.