deep-agents-orchestration

Workflow orchestration with human-in-the-loop approvals is handled by configuring interrupts for specific tools so actions can be paused, edited, or rejected. This requires a checkpointer to save the execution state and resume the flow after human review.

Subagent delegation is the process of spawning stateless subagents using a task tool to handle specialized, isolated work. It allows a primary agent to assign focused tasks, such as research, to ephemeral subagents with specialized prompts and tools.

To persist todo lists across multiple invocations, use the write_todos function with a thread_id. This automatically creates, updates, and tracks task statuses to maintain a persistent plan for multi-step workflows across different sessions.

Yes, a checkpointer is required to pause agent execution for human approval. It saves the workflow state at the interrupt point, allowing the agent to resume execution correctly once the human approves, edits, or rejects the pending action.

Yes, you can require human approval before running high-risk tools by configuring interrupts for specific actions. This gated execution mechanism prevents unsafe or unwanted operations by pausing the workflow until a human reviews the request.

Use workflow orchestration for subagent task delegation when coordinating multi-step AI workflows that need isolated specialist execution, session-persistent task tracking, and gated execution to prevent unsafe actions in complex environments.