The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/check-setup.py uses the __import__ function to verify the presence of required Python packages (autogen_agentchat, autogen_ext, autogen_core). The package names are hardcoded in the script and do not rely on external or user-provided input, making this a standard environment check.
[COMMAND_EXECUTION]: The skill documentation and templates (e.g., templates/code-execution.py) feature the use of LocalCommandLineCodeExecutor, which allows an agent to execute LLM-generated code on the host machine. The skill provides explicit security warnings, labeling this for 'Development Only' and 'trusted environments,' while recommending DockerCommandLineCodeExecutor for production to ensure containerized isolation.
[INDIRECT_PROMPT_INJECTION]: As a multi-agent framework, AutoGen agents ingest and process outputs from other agents which may contain instructions.
Ingestion points: Agents receive messages from other agents (e.g., AssistantAgent, UserProxyAgent) via conversation streams.
Boundary markers: The skill documents control mechanisms like is_termination_msg, max_turns, and MaxMessageTermination to limit conversation scope.
Capability inventory: Agents have access to local or Docker-based code execution as described in references/code-execution.md.
Sanitization: The skill promotes environment isolation (Docker) as the primary mitigation for executing untrusted agent-generated code.

autogen