The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The multi-agent architectures described (Supervisor, Swarm, and Hierarchical) rely on data passing between agents, creating a surface where untrusted input from one agent can affect the execution of another. If an agent responsible for gathering external data (e.g., a researcher) ingests malicious instructions, those instructions could be executed by the supervisor or aggregated into final outputs.
Ingestion points: SupervisorAgent.run_workflow in scripts/coordination.py and the researcher_node example in references/frameworks.md both handle potentially untrusted input during task execution.
Boundary markers: The provided implementation snippets do not include explicit boundary markers or delimiters (e.g., XML tags or specific 'ignore' instructions) to separate data from instructions.
Capability inventory: The architecture provides capabilities for persistent storage via FileSystemCoordination and inter-agent communication, which could be exploited if an injection occurs.
Sanitization: While the skill's text documentation mentions validation as a mitigation, the actual code examples do not demonstrate sanitization or filtering of external content before processing.

multi-agent-patterns