The Agent Skills Directory

Remote Script Execution: The SKILL.md file documents a CLI installation method using a curl | bash pattern from cli.genkit.dev. This is a common developer utility for environment setup and originates from a vendor-controlled domain.
Command Execution via Model Context Protocol (MCP): The documentation in references/genkit_mcp.md describes how to run external tools via npx and aggregate their capabilities. This feature allows agents to extend their functionality through external protocols.
Filesystem Middleware Access: As detailed in references/genkit_middleware.md and references/agents-artifacts.md, the framework includes middleware that provides agents with tools to list, read, and write files. These capabilities are designed to be restricted to a specific root directory for workspace-aware tasks.
Indirect Prompt Injection Surface: The agents and multi-agent systems described (e.g., in references/agents-multi-agent.md) process data from potentially untrusted sources, including user messages and tool outputs. This creates a standard attack surface where agents might be influenced by embedded instructions, which is a typical consideration for this class of application.
Dynamic Execution Patterns: The framework supports custom agent logic and remote action definitions (as seen in references/agents-custom.md and references/genkit.md) that involve orchestrating multiple model calls and executing code-defined workflows.

developing-genkit-dart