The Agent Skills Directory

[SAFE]: The skill consists of instructional markdown and Python code snippets for architecture design. It does not include executable scripts, external dependencies, or commands that interact with the system environment.
[INDIRECT_PROMPT_INJECTION]: The architectural patterns described (such as ReAct and Orchestrator-Workers) naturally involve the processing of untrusted user input or LLM-generated outputs. These represent a vulnerability surface inherent to the agent designs being taught.
Ingestion points: Untrusted data enters via user_input in the route_request function, the task parameter in Orchestrator.run, and the question variable in the react_agent pattern in SKILL.md.
Boundary markers: The example prompt templates (e.g., REACT_PROMPT) are simplified for educational clarity and do not include explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The patterns demonstrate tool execution capabilities via the execute_tool methods and multiple sequential LLM calls.
Sanitization: As a high-level architectural guide, the code snippets do not include implementation-specific sanitization or input validation logic.

agents