The complexity of multi-agent collaboration stems not only from the limitations of large models but also from the crucial importance of better controlling session identification, routing, and lineage relationships among agents.

This presentation aims to share how OpenClaw uses session control to separate the management and execution of multi-agent tasks, introducing the concept of a Team. Each team task is modeled as a room, consisting of a command session and multiple execution sub-agent sessions.

Attendees will learn practical design patterns for session-based agent orchestration from OpenClaw, crucial for any agentic orchestration implementation.