The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents a multi-agent interaction framework, which creates a potential surface for indirect prompt injection as agents process data from external participants while possessing tool-using capabilities.
Ingestion points: AgentClient components receive and process EV_TEXT envelopes from other agents or human users through the central Hub (as described in SKILL.md).
Boundary markers: The quickstart implementation examples do not utilize specific delimiters or instructions for the agent to ignore potential commands embedded within inbound message text.
Capability inventory: Registered agents can have access to significant tools, including identity-level operations like delegate, peers, channels, and tasks via the default NetworkPlugin, as well as custom tools such as end_conversation.
Sanitization: No explicit sanitization or validation of the content within inter-agent envelopes is demonstrated in the provided code examples.
[SAFE]: The skill contains no malicious patterns, hardcoded credentials, obfuscated content, or unauthorized network operations. All demonstrated functionality is consistent with the stated purpose of building a multi-agent coordination network and follows standard practices for the vendor's framework.

ag2-network-quickstart