Building Multi-Agent Architectures with LiveKit Agents

Overview

When designing a multi-agent architecture with LiveKit, developers often ask:

• Is function_tool the best way to transfer control or tasks between agents?

• Is using a UserData class the best way to make agents stateless?

Based on internal development and customer implementations, the answer to both is yes.

Transferring Between Agents with function_tool

• function_tool is the recommended way for agents to invoke each other’s capabilities or pass control.

• It provides a structured, reliable mechanism for chaining agents without requiring complex manual orchestration.

• This ensures clear responsibility boundaries and makes agents more composable.

📌 Example use case:
An intake agent can forward structured data to a specialist agent (e.g., medical triage, billing, or scheduling) via function_tool, rather than embedding all logic into a single monolithic agent.

Stateless Agents with UserData

• By default, agents should remain stateless to maximize scalability and reusability.

• The UserData class provides a consistent way to attach contextual information across interactions without binding state to the agent itself.

• This makes it easier to:

  • Run agents across distributed systems.

  • Maintain clean separation between logic and session-specific data.

  • Support fault tolerance and recovery.

Reference Example

A working example of these concepts is available in the LiveKit Python agents examples repo.

This example demonstrates:

• Multiple agents collaborating to handle a medical office triage flow.

• Use of function_tool for inter-agent communication.

• Application of UserData for maintaining context without binding state to an agent.

4. Key Takeaways

• ✅ Use function_tool to delegate tasks between agents.

• ✅ Use UserData for context management, keeping agents stateless.

• ✅ Stateless, modular agents are easier to scale, maintain, and extend.

• ✅ Example implementations are available in the LiveKit Python agents repo.