The Agent Skills Directory

[PROMPT_INJECTION]: The skill enables inter-agent communication through sessions_spawn and sessions_send, which creates an inherent attack surface for indirect prompt injection. In this model, instructions or data passed between agents could potentially be used to manipulate the receiving agent's behavior.
Ingestion points: Identified in SKILL.md; the task parameter in sessions_spawn and the message parameter in sessions_send are used to pass data to other agents.
Boundary markers: The documentation does not specify the use of delimiters or "ignore" instructions to separate untrusted inter-agent data from system prompts.
Capability inventory: Agents have the capability to launch new execution instances and perform multi-round negotiations.
Sanitization: The framework documentation does not detail specific input validation or sanitization for messages exchanged between agents.
[COMMAND_EXECUTION]: The sessions_spawn tool provides a mechanism for agents to launch new execution environments. The documentation notes that in run mode, these processes are invisible to the user. While this is a functional requirement for background tasks, it represents a high-capability feature that relies on the system's built-in depth limits and whitelists for safety.

multi-agent-communication