The Agent Skills Directory

[DYNAMIC_EXECUTION]: The utility scripts designed to analyze and visualize graph structures perform dynamic loading and execution of local Python modules.
Evidence: In scripts/validate_agent_graph.py and scripts/visualize_graph.py, the load_graph_from_path function uses importlib.util.module_from_spec and spec.loader.exec_module(module) to load user-supplied Python files.
Context: This behavior is intended to allow the tools to inspect the compiled graph object within the user's project, but it means any code in the targeted file will be executed in the current environment.
[INDIRECT_PROMPT_INJECTION]: The multi-agent patterns described in the skill are designed to ingest untrusted data (user messages) and interpolate them into prompts for LLM agents.
Ingestion points: Data enters the system via the messages list defined in the TypedDict state schemas in SKILL.md and references/state-management-patterns.md.
Boundary markers: The implementation guides use system messages to define agent roles but do not include explicit delimiters or "ignore instructions" guards when interpolating user content.
Capability inventory: The agents are granted capabilities to perform web searches (e.g., via Tavily), write content, and coordinate other agents.
Sanitization: No sanitization or input validation logic is present in the provided templates to mitigate adversarial input embedded in processed data.

langgraph-agent-patterns