说实话，咱就是说A2A协议这个东西，你要是搞AI智Neng体开发，不了解还真不行。

一、MCP:上下文协议的标准化尝试.{ name : code-review-agent , description : 代码审查助手,擅长发现潜在bug和优化建议 , url : https://agent.example.com/a2a , version : 1.0.0 , capabilities :{ streaming :true, pushNotifications :false, stateTransition :true}, skills :, "outputModes": , } ],}@dataclassclass Task: id: str status: str = "submitted" artifacts: list = field messages: list = fieldtasks: dict = {}def handle_jsonrpc -> dict: """处理 JSON-RPC 请求""" method = request.get params = request.get req_id = request.get if method == "message/send": task_id = params.get.get+}") user_text = "" for part in params.get.get: if part.get == "text": user_text = part task = Task tasks = task # 模拟翻译 translated = f" {user_text}" task.status = "completed" task.artifacts = } ] return { "jsonrpc": "", "id": req_id, "result": asdict, } return { "jsonrpc": "", "id": req_id, "error": {"code": -, "message": f"Method not found: {method}"}, }# ============================================================# 第二部分：A2A Client# ============================================================async def discover_agent -> dict: """第一步：读取 Agent Card，了解对方NengZuo什么""" async with httpx.AsyncClient as client: resp = await client.get return resp.jsonasync def send_task -> dict: """第二步：发送 JSON-RPC 请求，委托翻译任务""" payload = { "jsonrpc": "", "id": , "method": "message/send", "params": { "message": { "role": "user", "parts": , } }, } async with httpx.AsyncClient as client: resp = await client.post return resp.jsonasync def demo_client: """完整客户端流程：发现 → 委托 → 获取结果""" base_url = # Step : 获取结果 # printif __name__ == "__main__ ": print

整个流程的关键点：Client 不需要知道 Server内部用的是什么模型、什么框架。它只需要读Agent Card 、发JSON-RPC请求、等Task完成。明天Server从Claude换成Gemini ，Client代码一行不用改。

参考来源： Agent-to-Agent Protocol ）是Agent之间的HTTP —— 一个标准化的通信协议，让任何框架、任何模型构建的AgentdouNeng互相发现、互相委托任务、互相交换结果。