The Agent Skills Directory

[Agent Orchestration Surface]: The skill documents how to delegate tasks to subagents using instructions that are dynamically passed into the task tool. This creates a surface area where instructions from one agent layer could influence another. However, this is inherent to the orchestration framework's design.
Ingestion points: The instruction parameter within the task tool in SKILL.md allows for passing natural language to subagents.
Boundary markers: The provided examples do not explicitly define boundary markers for input sanitization between agent tiers.
Capability inventory: The documentation references high-privilege tools such as write_file, execute_sql, and deploy_to_prod as examples of what agents might orchestrate.
Sanitization: No specific sanitization or escaping logic is included in the documentation snippets, as they focus on high-level orchestration patterns.
[Human-In-The-Loop Security Control]: A significant portion of the skill is dedicated to configuring HumanInTheLoopMiddleware. This provides a security boundary by requiring manual approval or allowing for manual edits before the agent executes high-risk tools like database writes or deployments, which serves as a primary mitigation for potential orchestration risks.

deep-agents-orchestration