A2A Sample Methods and JSON Responses

I have a question: with the agent.json in A2A being so complex, how do LLMs call Agents?