The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions include steps to install the ag2 Python package via pip and bootstrap a frontend using the copilotkit framework. These are legitimate development operations using established package managers and official vendor/service resources.
[PROMPT_INJECTION]: The skill architecture involves processing untrusted user input from a web endpoint through an AI agent, which is a common surface for indirect prompt injection. 1. Ingestion points: The /chat endpoint in the provided FastAPI example accepts user messages via the RunAgentInput model. 2. Boundary markers: Not explicitly defined in the minimal code recipe, though the documentation notes the need for input handling. 3. Capability inventory: The agent is configured to handle streaming text and tool call lifecycles. 4. Sanitization: The skill includes a specific production checklist advising developers to validate and authorize tool inputs server-side to mitigate potential injection risks.

ag2-ag-ui