일본 신입 변호사들 ‘나 떨고 있니'…글로벌 ‘법률 AI’ 잇단 상륙

The article discusses an architecture choice made by TaskHandoff to address three specific problems that arise as AI tasks become more numerous: task distribution across multiple machines, each task requiring its own independent and disposable environment, and long-running tasks not being killed when the browser is closed or the interface is disconnected.

The solution involves separating the control plane from the execution plane. The control plane is responsible for managing node lists, instance dashboards, AI session views, and chat credentials and message routing. The Node Agent manages node local configurations, runtime resources, directory lists, and instance lifecycles. Controlled instances contain the actual work area, applications, AI sessions, and triggers, and can run independently.

The key point is that instance lifecycles belong to the Node Agent, so stopping or restarting the control plane does not affect the instances running on the node. The AI session maintains its conversation state, while the control plane handles credentials, bindings, command parsing, and routing. This design prevents the control plane from stopping Docker instances when restarted, and allows AI sessions to continue running even if the browser is closed or the interface is disconnected.

The article provides a step-by-step explanation of how the system works, including how to use the browser/desktop/mobile interfaces, manage nodes and instances, and handle long-running tasks across multiple machines and devices.