Skip to main content
An agent is a configured conversation runner. It holds a system prompt, a voice, an LLM model, and a set of tools / playbooks / FAQs it can draw on. End-users talk to agents — never directly to the underlying model.

What’s on an agent

Lifecycle

Only active agents accept inbound calls and originate outbound calls.

Operations

Where agents talk

An agent is just config. The actual conversation runs in the Call Processor (Pipecat-based, GPU-friendly). You connect the agent to one of three transports:
  1. PSTN (phone) — assign a phone_number to the agent (see Telephony). Inbound rings the agent; outbound goes through POST /api/v1/calls/originate.
  2. WebRTC widget — embed the JS SDK on a customer page (see Widget).
  3. Direct WebSocket — for custom clients that speak the gateway protocol directly.