deep-agents-orchestration

To coordinate subagents and human approvals in agent workflows, use an orchestration layer that delegates tasks via a task tool and routes sensitive decisions through human approval interrupts to prevent unsafe operations.

Tracking multi-step task planning across agent sessions requires a planning middleware that uses TodoList to explicitly map and persist task states across interactions. This ensures continuous progress visibility for ongoing complex workflows.

Human in the loop HITL with stateless subagents works by using a checkpointer and thread_id persistence to pause execution. The main agent triggers an interrupt to route sensitive decisions to humans before resuming the stateless subagent.

You can delegate specialized work to subagents without bloating the main agent by using SubAgentMiddleware. It delegates deterministic tasks via the task tool, keeping the main agent lean while distributing specific workload execution.

You can delegate specialized work to subagents without bloating the main agent by using middleware to route tasks via the task tool. This delegates deterministic execution while keeping the main agent lean and focused on orchestration.

You should use a checkpointer and thread_id for agent orchestration when managing human approval interrupts. They provide the necessary state persistence to pause workflows safely and resume operations after human validation is completed.