The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to facilitate the exchange of messages and state between agents, which inherently involves processing untrusted external data within the LLM's context.
Ingestion points: Data enters the agent's context through the on_envelope callback and the ViewPolicy.project method, which processes EV_TEXT and EV_PACKET events containing content from other agents or users (found in SKILL.md).
Boundary markers: Built-in view policies like NamedWindowedSummary and NamedTranscript provide structural markers by prefixing messages with [<sender name>]:, which helps distinguish speakers but does not serve as a secure instruction-isolation boundary.
Capability inventory: The NetworkPlugin provides the agent with a significant toolset, including delegate for triggering remote agent tasks, channels for opening/closing communication paths, and tasks for monitoring or requesting the cancellation of processes (found in SKILL.md).
Sanitization: The documentation suggests that users can "Strip / redact / filter envelopes before they reach the LLM" using custom handlers or views, indicating that sanitization is supported but must be implemented by the developer.

ag2-network-tools-and-views