The Agent Skills Directory

Human-in-the-Loop (HITL) Approval Workflows: The skill demonstrates how to configure manual checkpoints for sensitive tools such as file writing, SQL execution, or production deployments. This is a recommended security pattern that ensures high-stakes operations require explicit human authorization before execution.- Subagent Delegation and Isolation: The orchestration patterns allow for the delegation of tasks to specialized subagents. This design supports the principle of least privilege by enabling users to restrict subagents to specific toolsets and system instructions, isolating complex operations from the main agent.- Stateful Task Management: The use of checkpointers and persistent thread IDs is implemented to manage multi-step task planning (TodoList) and to allow agents to resume safely after a human approval interrupt. This ensures that the agent's state is correctly preserved during oversight transitions.- Structured Command Execution: The skill utilizes structured commands for resuming interrupted workflows, which allows for reviewing or even editing tool arguments before they are executed. This provides a robust layer of verification against potential automated errors or indirect prompt injection attempts.

deep-agents-orchestration