Add a new load_skill tool to the agent graph so any agent/sub-agent can dynamically load a specific skill.md (or skill bundle) at runtime when needed.
This lets agents pull in the exact guidance for a tool, vulnerability type, or workflow, right before acting