The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of autonomous agent teams that communicate via natural language, which is an inherent surface for indirect prompt injection if processing untrusted data.
Ingestion points: Defined in SKILL.md and templates via initiate_chat and team.run inputs.
Boundary markers: Encourages the use of system_message and is_termination_msg to define agent boundaries.
Capability inventory: The framework supports shell command execution and tool calling through LocalCommandLineCodeExecutor, DockerCommandLineCodeExecutor, and register_function.
Sanitization: The skill explicitly promotes the use of DockerCommandLineCodeExecutor to isolate the environment from potentially malicious generated code.
[DYNAMIC_EXECUTION]: A setup utility script uses dynamic module loading to verify the installation environment.
Evidence: scripts/check-setup.py calls __import__() to check for the presence of the autogen_agentchat and autogen_ext libraries.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing official framework packages and tools from well-known sources.
Evidence: Mentions installation of autogen-agentchat via pip and fetching the Playwright MCP server using npx @playwright/mcp@latest in references/tool-integration.md.

autogen