The Agent Skills Directory

[COMMAND_EXECUTION]: The library provides a built-in execute tool, documented in references/tools.md, which allows agents to run arbitrary shell commands on the host system. This tool is enabled when using the FilesystemBackend as shown in references/patterns.md.
[EXTERNAL_DOWNLOADS]: The documentation in SKILL.md includes examples of integrating Model Context Protocol (MCP) servers using npx to fetch and run packages from the @modelcontextprotocol organization.
[PROMPT_INJECTION]: The skill facilitates the creation of agents that combine high-privilege capabilities with data ingestion from potentially untrusted sources, creating an indirect prompt injection surface.
Ingestion points: Agents can ingest external data through tools such as web_search (SKILL.md) and read_file (references/tools.md).
Boundary markers: The provided implementation examples and patterns do not include the use of delimiters or specific instructions to ignore embedded commands in processed data.
Capability inventory: Agents built with this library have access to powerful tools including execute (shell commands), write_file, and edit_file, as listed in the tools reference.
Sanitization: The documentation does not describe or suggest mechanisms for input validation or content sanitization before data is processed by the agent.

deepagents-implementation