The Agent Skills Directory

File System Interaction Capabilities: The skill documents the use of built-in tools for file management, including read_file, write_file, and edit_file. These capabilities allow agents to interact with the filesystem to manage context and persistent data as part of their intended architectural design.
Indirect Prompt Injection Surface: The agent framework described is designed to ingest and process data from various external sources, which represents a potential surface for indirect instructions.
Ingestion points: User messages provided via agent.invoke(), data retrieved through filesystem tools (read_file, glob, grep), and documentation loaded from skill directories (SKILL.md).
Boundary markers: The documentation does not explicitly show specific delimiter strategies for external content, relying instead on the framework's internal middleware and model guardrails.
Capability inventory: The framework includes tools for file system modification (write_file, edit_file) and a delegation mechanism (task tool) to spawn specialized subagents.
Sanitization: Standard sanitization or filtering logic is not detailed in these high-level configuration examples, as it is typically managed by the underlying LLM provider and framework middleware layers.

deep-agents-core