The Agent and AgentSession classes serve different but complementary roles in the LiveKit Agents framework: Agent Class: • Represents the AI logic and behavior of your voice agent. • Defines the personality, instructions, and capabilities of your AI assistant. • Can include custom prompts, tools, and other logic. • Multiple agents can exist within a single session for workflows and handoffs. AgentSession Class: • Manages the runtime infrastructure for the agent. • Handles media streams, speech/LLM components, and tool orchestration. • Coordinates components like STT, TTS, VAD, and LLM interactions. • Manages turn detection, interruptions, and technical aspects of the conversation. Relationship: • The

AgentSession

acts as a container that runs one or more

Agent

instances. • Each session requires at least one

Agent

to orchestrate. • Typically, you create an

Agent

(defining AI behavior), then an

AgentSession

(setting up the technical infrastructure), and start the session with the agent. This separation allows you to focus on the AI logic in the

Agent

class while letting

AgentSession

handle the complex real-time media processing and coordination. Sources: livekit-agents/livekit/agents/__init__.py | Building voice agents | LiveKit Docs | Workflows | LiveKit Docs | Agents Core Architecture | livekit/livekit_composite | DeepWiki