互联网之所以能成为互联网，是因为所有计算机都说同一种"话"——TCP/IP。
Agent世界也在等待它的TCP/IP时刻。6月，这个时刻似乎更近了一步。
从Google开源到Linux基金会
Google正式将A2A（Agent-to-Agent）协议的开源仓库移交Linux基金会治理，最新版本v0.3.0已经具备了生产级协议的核心要素：基于JSON-RPC 2.0 over HTTP(S)的通信层、通过AgentCard实现的能力自描述与服务发现、Task生命周期管理（submitted/working/completed/failed）、TextPart/FilePart/DataPart三部分消息协议。
更重要的是官方SDK的覆盖：Python、JavaScript、Java、.NET四套官方SDK同时发布，意味着主流技术栈的开发者都可以直接接入，不需要自己从头实现协议细节。同步、异步、流式（SSE）三种交互模式也全部支持——从简单的问答到复杂的长周期协作任务都能覆盖。
阿里云AgentRun：生产环境的第一道考题
协议有了，真正难的是生产环境的工程问题。阿里云在6月16日发布的AgentRun平台，正面回答了这些问题。
多Agent协作在Demo里看起来很美——几个Agent聊聊天就把事办了。但到了生产环境，问题立刻来了：Agent怎么找到彼此？（注册中心/服务发现）跨Agent的调用怎么做鉴权？（API Key/HTTP Basic Auth）不同环境（开发/预发/生产）怎么隔离？（Workspace逻辑隔离）调用链出了问题怎么排查？（分布式链路追踪）
AgentRun给出了一套相对完整的解法：通过Discovery Endpoint按环境管理发现入口，内置凭证体系处理跨Agent鉴权，Workspace实现环境隔离。在"希希咖啡厅"的演示场景中，coffee_agent负责点单、delivery_agent负责配送，一个超级Agent（Orchestrator）负责意图拆解和子任务调度——整个流程通过A2A协议协同完成。
微信率先接入，荣耀第一个落地
最有信号意义的消息是：微信已经基于A2A协议向手机厂商开放Agent能力，荣耀成为第一个落地合作方。
这意味着什么？意味着你的手机系统助手（比如荣耀的YOYO）未来可以直接调用微信里的Agent能力——你不需要打开微信，对着手机说"帮我给张三发个消息说我晚到十分钟"，系统Agent通过A2A协议和微信Agent通信，直接完成任务。
"A2A之于Agent，就像HTTP之于Web。协议本身不产生价值，协议之上的生态才产生价值。"
—— 云原生技术社区评论
A2A和MCP的分工正在清晰
很多人在问：已经有了MCP（Model Context Protocol），为什么还需要A2A？两者的定位正在变得清晰：
MCP解决的是"模型连接工具"的问题——Agent怎么调用数据库、怎么读文件、怎么操作API，它是单向的、无状态的、事务式的，就像浏览器和服务器之间的请求/响应。
A2A解决的是"Agent连接Agent"的问题——两个对等的Agent怎么协商任务、怎么分配工作、怎么同步状态，它是双向的、有状态的、协作式的，就像两个人之间的对话和分工。
两者不是竞争关系，而是互补关系。一个Agent通过MCP调用工具完成具体操作，通过A2A和其他Agent协作完成复杂任务。MCP是Agent的"手"，A2A是Agent的"嘴"。
协议的地基正在一层层打好。当Agent之间可以无障碍对话、协作、分工的时候，真正的"Agent互联网"才会从PPT走进现实。
明天见。
The internet became the internet because all computers spoke the same "language" -- TCP/IP.
The Agent world has been waiting for its TCP/IP moment. In June, that moment seemed to draw a step closer.
From Google Open Source to Linux Foundation
Google formally transferred the open-source repository for the A2A (Agent-to-Agent) protocol to Linux Foundation governance. The latest version v0.3.0 already has the core elements of a production-grade protocol: a JSON-RPC 2.0 over HTTP(S) communication layer, capability self-description and service discovery via AgentCard, Task lifecycle management (submitted/working/completed/failed), and a three-part message protocol of TextPart/FilePart/DataPart.
More importantly, official SDK coverage: four official SDKs released simultaneously -- Python, JavaScript, Java, .NET -- meaning developers across mainstream stacks can directly integrate without implementing protocol details from scratch. Sync, async, and streaming (SSE) interaction modes are all supported -- covering everything from simple Q&A to complex long-cycle collaboration tasks.
Alibaba Cloud AgentRun: The First Production Exam
Protocols exist; the truly hard part is production engineering. AgentRun, launched by Alibaba Cloud on June 16, directly answers those questions.
Multi-Agent collaboration looks beautiful in demos -- a few Agents chat and get things done. But in production, problems immediately arise: how do Agents find each other? (Registry/service discovery) How do you authenticate cross-Agent calls? (API Key/HTTP Basic Auth) How do you isolate environments (dev/staging/prod)? (Workspace logical isolation) How do you troubleshoot broken call chains? (Distributed tracing)
AgentRun provides a relatively complete solution: managing discovery endpoints by environment via a Discovery Endpoint, built-in credential systems for cross-Agent authentication, Workspace for environment isolation. In the "Xixi Cafe" demo scenario, coffee_agent handles ordering, delivery_agent handles delivery, and a super-Agent (Orchestrator) handles intent decomposition and subtask scheduling -- the entire flow is coordinated through the A2A protocol.
WeChat Connects First; Honor Is First Implementation
The most signal-worthy news: WeChat has already opened Agent capabilities to phone manufacturers based on the A2A protocol, with Honor as the first implementation partner.
What does this mean? It means your phone's system assistant (like Honor's YOYO) can directly invoke Agent capabilities within WeChat in the future -- you don't need to open WeChat; just say to your phone "send Zhang San a message saying I'll be ten minutes late," and the system Agent communicates with WeChat's Agent via A2A to complete the task directly.
"A2A is to Agents what HTTP was to the Web. Protocols themselves don't create value; the ecosystems on top of them do."
-- Cloud-native community commentary
The Division of Labor Between A2A and MCP Is Clarifying
Many people ask: we already have MCP (Model Context Protocol); why do we need A2A? The positioning of the two is becoming clear:
MCP solves the "model connecting to tools" problem -- how Agents call databases, read files, operate APIs. It's one-way, stateless, transactional, like request/response between browser and server.
A2A solves the "Agent connecting to Agent" problem -- how two peer Agents negotiate tasks, allocate work, synchronize state. It's bidirectional, stateful, collaborative, like dialogue and division of labor between two people.
The two aren't competitive; they're complementary. An Agent uses MCP to call tools for specific operations, and A2A to collaborate with other Agents on complex tasks. MCP is the Agent's "hands"; A2A is the Agent's "mouth."
The protocol foundation is being laid layer by layer. When Agents can converse, collaborate, and divide labor without friction, the true "Agent internet" will move from PPT into reality.
See you tomorrow.
A2A之于Agent，就像HTTP之于Web。协议本身不产生价值，协议之上的生态才产生价值。
—— 云原生技术社区评论
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 5 个源信号生成，经编辑部人工审核。
Generated by Dawn Vision's cognitive engine from 5 source signals, editorially reviewed.
