You're asking how

session.update_agent

differs from returning an agent from a tool — here’s a concise comparison and when to use each. • Initiator ◦ Returning an agent from a tool: initiated by the LLM during tool execution (model-driven). ◦ `session.update_agent`: initiated by developer code (programmatic). • Flow ◦ Returning from a tool: tool runs, returns a new Agent (or handoff object); framework performs the handoff after the tool completes and any LLM reply. ◦ `session.update_agent`: code calls

session.update_agent(new_agent)

and the handoff happens immediately without a tool or LLM step. • Use cases ◦ Returning from a tool: use when the LLM should decide to route the conversation (dynamic, context-based routing). ◦ `session.update_agent`: use when your application logic or external events determine the handoff (state, timers, event handlers). • Platform/availability ◦ Returning from a tool: supported in both Node.js (via

llm.handoff()

) and Python. ◦ `session.update_agent`: available in Python only. • Side effects ◦ Both result in a new active agent and room-wide sync; returning from a tool also emits a

conversation_item_added

AgentHandoff item as part of the chat history. Which do you want to implement? I can show a short Python example for

session.update_agent

or a tool-return handoff example depending on your use case. Sources: Agents and handoffs | LiveKit docs | Agent session | LiveKit docs | Agents (Python) | livekit/livekit_composite | DeepWiki