ag2-network-tools-and-views

AG2 Network — Tools, Views & Custom Handlers

Everything on the agent/client side of the network — the mirror image of ag2-network-governance (which is hub-side). This covers what the LLM sees of the channel (views), which actions it can take (the five auto-injected plugin tools plus the adapter-owned say), what other agents know about it (skill markdown), how to replace the handler entirely, and the full Envelope reference.

Prerequisite: read ag2-network-quickstart first. This skill assumes you know Hub.open, HubClient.register, the channel lifecycle, and basic agent_client.open(...) / channel.send(...).

When to use

• "Limit / extend the LLM's network tool surface" (plugin tools vs. adapter tools_for)
• "Write a custom envelope handler"
• "Build a gateway / headless worker that doesn't run an LLM"
• "Add a non-LLM participant — a person at a UI, a bridge, a scripted seeder" (HumanClient)
• "Customise what each agent sees of the channel history (view policy)"
• "Strip / redact / filter envelopes before they reach the LLM"
• "Wire peer discovery via skill markdown"
• "Send a custom event type (myapp.review_request, …)"
• "I need the EV_* constants list / Envelope shape / audience semantics"

Network tools — plugin tools vs. adapter tools

An agent's per-turn LLM tool list is assembled from two streams:

1. Plugin tools — when you register with the default attach_plugin=True, NetworkPlugin adds five identity-level, channel-agnostic verbs to agent.tools: peers / channels / tasks / context / delegate. Same behaviour in any channel.
2. Adapter tools — the channel's adapter offers channel-specific tools per turn via adapter.tools_for(client, metadata, state, participant_id); the default handler resolves them and merges them into agent.ask(tools=...) (cached per (adapter, client) so the schema build cost is paid once). The only built-in adapter tool is say.

So say is not a plugin tool — it's offered by the adapter, gated by turn state: