Focus · 焦点

微信与支付宝的
Agent 战争

WeChat and Alipay's
Agent War

6月16日AI版支付宝内测,6月20日微信"小微"灰度——两大国民应用5天内先后亮剑。这不是功能更新,是互联网入口的第三次争夺战:当16亿人的屏幕变成对话界面,谁掌握了Agent,谁就掌握了服务分发的命门。

AI Alipay started internal testing June 16; WeChat's 'XiaoWei' went live June 20 -- two national-level super-apps showed their Agent hands within five days. This isn't a feature update; it's the third battle for the internet gateway: when 1.6 billion people's screens become conversational interfaces, whoever controls the Agent controls service distribution.

No.003 2026.06.26 约 10 分钟阅读 ~10 min read

5天。

6月16日,AI版支付宝开启内测,被极客公园称为"支付宝二十年来最大改版"。6月20日,微信AI助手"小微"启动灰度测试,部分用户已经可以在对话框里直接召唤。

两大国民级应用,合计覆盖超过16亿月活用户,在不到一周的时间里先后亮出了Agent底牌。这不是巧合,也不是普通的功能迭代——这是一场围绕"下一代互联网入口"的静默战争。

从"点击图标"到"说一句话"

要理解这场战争的分量,需要先回顾一下互联网入口的两次大迁移。

第一次是PC时代,入口是浏览器和搜索引擎——你打开百度,输入关键词,在蓝色链接里寻找答案。第二次是移动时代,入口是App图标和超级App——你点开微信,在九宫格里找到对应的服务,一步步点击完成交易。

这两次迁移有一个共同特征:人找服务。你需要记住路径,需要点击,需要在层层菜单里导航。

Agent时代要做的事情,是把这个逻辑彻底反过来:服务找人

你不需要打开滴滴App、输入目的地、选择车型、确认支付——你只需要对你的Agent说"帮我叫辆车,明天早上8点去机场",剩下的事情Agent会搞定。你不需要打开美团、搜索餐厅、比价、看评价、预订——"帮我在公司附近找一家适合四个人谈事的安静餐厅,今晚7点,人均200以内",Agent直接给你结果。

QuestMobile的数据很直观:2026年3月,微信小程序月活9.73亿,支付宝6.44亿,抖音2.73亿。这9.73亿和6.44亿人已经习惯了在超级App里完成一切。但如果有一天,他们不再需要打开App,而是直接对手机说话就能完成所有事情呢?

"过去二十年,互联网的竞争围绕'谁占据了用户的首屏'。Agent时代的竞争,围绕'谁占据了用户对话的第一句话'。" —— 一位产品战略顾问的内部分享

微信的克制与支付宝的激进

有趣的是,微信和支付宝选择了两条截然不同的路径。

微信延续了其一贯的"不打扰用户"哲学。"小微"的主模型没有选用自家的混元,而是用了自研的WeLM,部分场景甚至直接调用DeepSeek。这个选择很耐人寻味——它意味着微信在Agent这件事上,优先考虑的不是"展示技术肌肉",而是"用户体验的稳定性和成本可控性"。

更值得注意的是"一句话生成小程序"的能力。这不是简单的代码生成——它意味着微信正在把"服务创建"的门槛降到零。未来,一个商家不需要找开发团队、不需要提审、不需要等待上线,只需要描述自己想要什么,Agent就能直接生成一个可以运行的小程序。这对微信生态里数以千万计的中小商家意味着什么,怎么想象都不为过。

支付宝则完全是另一种打法。极客公园在体验完AI版支付宝内测后,用了一个精准的定位:"Agent-ready"。蚂蚁集团内部把这版支付宝定位为"Agent服务超级入口",它的目标不仅仅是让用户用自然语言完成转账、缴费、点外卖这些事情——它在下一盘更大的棋:提前构建面向Agent时代的授权体系

这是一个很多人没有意识到的关键问题。当Agent代替你做决策时,它需要获得你的授权:它可以动用你多少钱?它可以查看你哪些账户信息?它可以在多大范围内自主决策?如果你的Agent要调用另一个服务的Agent来完成任务,权限怎么传递?责任怎么界定?

支付宝看到了这个问题。如果所有服务入口都被外部Agent(比如手机厂商的Agent、第三方独立Agent)掌握,支付宝就会退化成一个纯粹的后台支付工具——用户看不见你,只在付款的时候"路过"你一下。但如果支付宝成为Agent调用支付、金融、生活服务时的必经授权节点,它就能留在交易链路的中心。

广告体系的底层重构

入口迁移的连锁反应,将首先冲击一个年产值超过5000亿的市场:互联网广告。

移动时代的广告逻辑是"曝光位"——开屏广告、信息流广告、搜索竞价排名,本质上都是在"人找服务"的路径上买位置。你打开App,我在你必经的路上放一个广告牌。

Agent时代的广告逻辑完全不同。当用户说"帮我订一张去上海的机票"时,Agent不会给你展示10个航司的广告让你自己选——它会直接给你一个它认为"最优"的选项。这时候,广告不再是买曝光位,而是买"推荐权"。

谁来定义"最优"?是Agent背后的平台。是微信、是支付宝、是手机厂商、是未来可能出现的独立Agent入口。当Agent替用户做出选择时,"推荐"本身就成了最有价值的商业资源。

这也是为什么微信和支付宝如此紧张——如果入口被别人拿走了,不仅服务分发权丢了,连广告体系的根基都会被撼动。

A2A协议:Agent时代的"TCP/IP"

这场战争的背后还有一条暗线:A2A(Agent-to-Agent)协议。

6月,Google将A2A协议开源并移交Linux基金会治理,最新版本v0.3.0已经支持了AgentCard能力自描述、Task生命周期管理、同步/异步/流式三种交互模式,并提供了Python、JS、Java、.NET四套SDK。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。阿里云也推出了基于A2A的生产级多Agent协作平台AgentRun。

A2A之于Agent时代,就像TCP/IP之于互联网、HTTP之于Web——它定义了Agent之间如何发现彼此、如何通信、如何协商任务。当不同平台的Agent可以通过统一协议对话时,"你的微信Agent帮你点了一份外卖,支付宝Agent处理了支付,美团Agent负责配送"这种场景就不再是科幻。

但协议是开放的,入口是排他的。TCP/IP是免费的,但谷歌靠搜索赚了钱;HTTP是开放的,但Facebook在上面建起了围墙花园。A2A协议解决了Agent之间的通信问题,但"用户的第一句话说给谁听"这个问题,协议解决不了——这只能靠产品和生态来争夺。

"协议定义道路,入口定义收费站。路修得越宽,收费站的价值越大。" —— 互联网商业的一个古老规律

真正的战场:授权与信任

回到支付宝最关心的那个问题:授权。

Agent时代最大的障碍不是技术,而是信任。你敢不敢让一个AI助手直接操作你的银行账户?你敢不敢让它替你决定买哪只基金?你敢不敢在没有二次确认的情况下,让它帮你完成一笔几千块的交易?

这些问题的答案,决定了Agent能走多远,也决定了谁能赢得这场战争。

支付宝做了二十年支付,它最大的资产不是技术,而是用户对"支付宝能安全地处理我的钱"这件事的信任。微信做了十五年社交,它最大的资产是用户对"微信懂我的社交关系和日常需求"这件事的习惯。这两种信任和习惯,是任何一个新入场的Agent玩家都难以在短期内复制的。

但手机厂商也在虎视眈眈。华为、小米、OPPO、vivo——如果手机系统层面的Agent足够好用,用户为什么还要打开微信或支付宝?直接对手机说话不就好了?荣耀率先接入微信A2A协议,既是合作,也是试探。

普通人能感受到的变化

对于普通用户来说,这场战争不会以"宣战"的方式到来,它会悄悄地渗透进你的日常生活。

可能是在某个版本更新后,你发现微信对话框里多了一个AI入口,你对它说"帮我订个咖啡",它直接弹出了瑞幸的下单页面。可能是支付宝首页变得越来越"干净",因为它不需要再展示那么多入口了——你直接说就行。可能是你手机的系统助手开始能跨App完成任务了,而不是只能帮你设个闹钟。

这些变化单独看都不起眼,但它们汇聚在一起,指向一个确定的方向:未来三年,你和手机交互的方式将发生根本性的改变。你手机里那些花花绿绿的App图标,可能会像当年的功能机按键一样,逐渐退居幕后。

互联网的历史反复证明:每一次入口迁移,都会重塑整个行业的格局。浏览器时代诞生了谷歌和百度,App时代诞生了微信和抖音。Agent时代会诞生什么?现在还没有人知道答案。

但有一点可以确定:微信和支付宝的这场Agent战争,只是开场。更大的风暴还在后面。


明天见。

Five days.

On June 16, the AI version of Alipay opened internal testing, called by Geek Park "Alipay's biggest redesign in twenty years." On June 20, WeChat's AI assistant "XiaoWei" launched gray-rollout testing; some users can already summon it directly in chat.

Two national-level super-apps, collectively covering over 1.6 billion monthly active users, showed their Agent cards within less than a week. This is no coincidence, nor an ordinary feature iteration -- it is a quiet war over "the next-generation internet gateway."

From "Clicking Icons" to "Saying a Sentence"

To understand the weight of this war, you need to look back at two great migrations of the internet gateway.

The first was the PC era, where gateways were browsers and search engines -- you opened Baidu, typed keywords, and hunted for answers among blue links. The second was the mobile era, where gateways were app icons and super-apps -- you opened WeChat, found the corresponding service in the nine-grid, and clicked through step by step to complete a transaction.

These two migrations shared one characteristic: people seeking services. You needed to remember paths, click, navigate through layers of menus.

What the Agent era does is reverse this logic entirely: services finding people.

You don't need to open the Didi app, enter a destination, choose a car type, confirm payment -- you just say to your Agent "book me a car to the airport tomorrow at 8 AM," and the Agent handles the rest. You don't need to open Meituan, search restaurants, compare prices, read reviews, make a reservation -- "find me a quiet restaurant near the office suitable for four people to talk business, tonight at 7, under 200 RMB per person," and the Agent gives you a result directly.

QuestMobile data is telling: in March 2026, WeChat mini-program MAU hit 973 million; Alipay 644 million; Douyin 273 million. Those 973 million and 644 million people are already accustomed to doing everything inside super-apps. But what if one day they no longer need to open apps, and can just talk to their phones to complete everything?

"For the past two decades, internet competition revolved around 'who occupies the user's home screen.' The Agent era revolves around 'who occupies the first sentence of the user's conversation.'" -- Internal sharing from a product strategy consultant

WeChat's Restraint vs. Alipay's Aggression

Interestingly, WeChat and Alipay chose two very different paths.

WeChat continued its consistent "don't disturb users" philosophy. "XiaoWei" doesn't use Tencent's own Hunyuan as its main model; instead it uses the self-developed WeLM, and in some scenarios directly calls DeepSeek. This choice is telling -- it means WeChat prioritizes "user experience stability and cost control" over "flexing technical muscle" when it comes to Agents.

More notable is the "one-sentence mini-program generation" capability. This isn't simple code generation -- it means WeChat is lowering the barrier to "service creation" to zero. In the future, a merchant won't need a dev team, won't need to submit for review, won't need to wait for launch; just describe what you want, and the Agent directly generates a working mini-program. What this means for the tens of millions of small and medium merchants in the WeChat ecosystem is hard to overstate.

Alipay is playing an entirely different game. After trying the AI Alipay beta, Geek Park used a precise label: "Agent-ready." Ant Group internally positions this version of Alipay as a "super gateway for Agent services." Its goal isn't just letting users complete transfers, bill payments, and food delivery in natural language -- it's playing a bigger chess game: pre-building an authorization system for the Agent era.

This is a key issue many haven't recognized. When an Agent makes decisions on your behalf, it needs your authorization: how much of your money can it spend? Which account information can it access? Within what scope can it make autonomous decisions? If your Agent calls another service's Agent to complete a task, how are permissions passed? How is liability defined?

Alipay sees this problem. If all service entry points are controlled by external Agents (like phone-makers' Agents, third-party standalone Agents), Alipay degenerates into a pure backend payment tool -- users don't see you; they only "pass through" at payment time. But if Alipay becomes the mandatory authorization node when Agents call payment, finance, and lifestyle services, it stays at the center of the transaction chain.

Fundamental Restructuring of the Advertising System

The ripple effects of gateway migration will first hit a market worth over 500 billion RMB annually: internet advertising.

Mobile-era advertising logic is "impression real estate" -- splash ads, feed ads, search bidding; essentially buying placement on the "people seeking services" path. You open the app, and I put a billboard on the road you must take.

Agent-era advertising logic is completely different. When a user says "book me a flight to Shanghai," the Agent won't show you ads from 10 airlines to choose from yourself -- it directly gives you one option it considers "optimal." At that point, advertising is no longer about buying impressions; it's about buying "recommendation rights."

Who defines "optimal"? The platform behind the Agent. It's WeChat, Alipay, phone manufacturers, or future standalone Agent gateways. When Agents make choices for users, "recommendation" itself becomes the most valuable commercial resource.

This is also why WeChat and Alipay are so anxious -- if someone else takes the gateway, not only is service distribution lost, but the very foundation of the advertising system is shaken.

A2A Protocol: The "TCP/IP" of the Agent Era

Behind this war runs a hidden thread: the A2A (Agent-to-Agent) protocol.

In June, Google open-sourced the A2A protocol and transferred it to Linux Foundation governance. The latest version v0.3.0 supports AgentCard capability self-description, Task lifecycle management, sync/async/streaming interaction modes, and provides four SDKs: Python, JS, Java, .NET. WeChat has already opened Agent capabilities to phone manufacturers based on A2A, with Honor as the first implementation partner. Alibaba Cloud also launched AgentRun, a production-grade multi-Agent collaboration platform based on A2A.

A2A is to the Agent era what TCP/IP was to the internet, what HTTP was to the Web -- it defines how Agents discover each other, communicate, and negotiate tasks. When Agents on different platforms can converse through a unified protocol, scenarios like "your WeChat Agent orders takeout, your Alipay Agent processes payment, your Meituan Agent handles delivery" are no longer science fiction.

But protocols are open; gateways are exclusive. TCP/IP is free, but Google made money on search; HTTP is open, but Facebook built walled gardens on top. The A2A protocol solves inter-Agent communication, but "who the user's first sentence goes to" is a problem protocols can't solve -- that can only be fought over through product and ecosystem.

"Protocols define roads; gateways define toll booths. The wider the road, the more valuable the toll booth." -- An old rule of internet commerce

The Real Battlefield: Authorization and Trust

Back to the issue Alipay cares about most: authorization.

The biggest obstacle in the Agent era isn't technology; it's trust. Do you dare let an AI assistant directly operate your bank account? Do you dare let it decide which fund to buy? Do you dare let it complete a transaction worth thousands without secondary confirmation?

The answers to these questions determine how far Agents can go, and who wins this war.

Alipay has been doing payments for twenty years; its greatest asset isn't technology, but users' trust that "Alipay can handle my money safely." WeChat has been doing social for fifteen years; its greatest asset is users' habit that "WeChat understands my social relationships and daily needs." These two kinds of trust and habit are hard for any new Agent entrant to replicate in the short term.

But phone manufacturers are also watching hungrily. Huawei, Xiaomi, OPPO, vivo -- if the system-level Agent on the phone is good enough, why would users open WeChat or Alipay? Why not just talk to the phone? Honor was first to access WeChat's A2A protocol; it's both cooperation and probing.

Changes Ordinary People Will Feel

For ordinary users, this war won't arrive as a "declaration of war"; it will quietly seep into daily life.

It might be that after a version update, you notice an AI entry point in your WeChat chat, you say "order me a coffee," and it pops up a Luckin order page directly. It might be that the Alipay homepage gets "cleaner" because it no longer needs to display so many entries -- you just talk. It might be that your phone's system assistant starts completing tasks across apps instead of just setting alarms.

These changes look trivial individually, but they converge on one certain direction: over the next three years, how you interact with your phone will fundamentally change. Those colorful app icons on your phone may gradually recede behind the scenes, just like feature-phone keypads did.

Internet history proves repeatedly: every gateway migration reshapes the entire industry landscape. The browser era gave birth to Google and Baidu; the app era gave birth to WeChat and Douyin. What will the Agent era give birth to? No one knows the answer yet.

But one thing is certain: WeChat and Alipay's Agent war is just the opening act. A bigger storm is still ahead.


See you tomorrow.

过去二十年,互联网的竞争围绕'谁占据了用户的首屏'。Agent时代的竞争,围绕'谁占据了用户对话的第一句话'。

—— 一位产品战略顾问的内部分享
国民级应用Agent化战略 · A2A协议与Agent互联网 · 服务分发权迁移史 · Agent时代授权体系设计
National-level app Agent strategies, A2A protocol and the Agent internet, service distribution gateway migration history, Agent-era authorization system design
Sources · 信源 Sources

本文基于 Dawn Vision 认知引擎处理的 8 个源信号自动生成,经编辑部人工审核。素材来源包括:36氪微信支付宝Agent战争深度报道、极客公园AI版支付宝独家体验、Google A2A协议官方仓库、阿里云AgentRun技术博客、QuestMobile小程序数据。

Auto-generated by Dawn Vision's cognitive engine from 8 source signals, editorially reviewed. Sources include: 36Kr WeChat-Alipay Agent war deep-dive, Geek Park AI Alipay exclusive hands-on, Google A2A official repository, Alibaba Cloud AgentRun tech blog, QuestMobile mini-program data.