Agent handoff in RealtimeModel works using the same tool-based workflow system as in the broader LiveKit Agents framework. Here’s how the process functions: • Tool-based Handoff: You trigger a handoff by returning an

Agent

instance from a tool function. This shifts conversation control to the new agent within the active session. • AgentSession Management: Handoffs use the

update_agent

method on the

AgentSession

, switching to another agent as needed. Example:

session.update_agent(SecondAgent(chat_ctx=session.chat_ctx))

• Function Tool Integration: With RealtimeModel, handoff logic is implemented inside a tool:

@function_tool()
async def transfer_to_specialist(self):
    """Transfer to specialist when complex issue detected"""
    return SpecialistAgent(chat_ctx=self.session.chat_ctx)

• Preserving Context: To maintain conversation history, pass

chat_ctx

from the current session to the new agent at handoff. • Session Usage: RealtimeModel instances are used directly in

AgentSession

, supporting both standard text-based and real-time workflows, including agent handoff mechanics via tools. Other features such as user data preservation and logic-based conditional handoffs are managed similarly to traditional LLM workflows. RealtimeModel’s built-in turn detection may impact the precise handoff timing, but not the workflow steps. For further architectural or code integration examples, let me know your specific use case or scenario details—comprehensive handoff workflows and advanced patterns are available for multiple-agent handoff and SIP/call transfer integration if needed. Sources: complex-agents/medical_office_triage/README.md | Agents 0.x migration guide | LiveKit Docs | Workflows | LiveKit Docs | Tool definition and use | LiveKit Docs | examples/avatar_agents/hedra/agent_worker.py | LLM Integrations | livekit/livekit_composite | DeepWiki