Reading through the documentation, I get a sense that the "agent" server could be any application that exposes an agent card and follows the protocol. Doesn't really have to be using LLM or be "agentic" (guess that's where the agent black box) comes in.