WeChat agent issues a "hero call", and half of the internet industry responds

The WeChat Agent is really coming.

The WeChat Open Platform has released content regarding the guidelines for WeChat AI developers.

The guidelines state that to provide users with a more intelligent interactive experience and help them discover and use mini - programs more efficiently, the WeChat Open Platform offers developers the ability to conveniently access the WeChat AI ecosystem while fully respecting their rights and autonomous choices.

After access, mini - programs will have the opportunity to be recommended and invoked by WeChat AI. Mini - programs that have not completed the access process cannot be invoked by WeChat AI.

The platform provides two access modes. In the automatic mode, the platform is authorized to read the mini - program source code during the submission review, without the need for additional development. In the development mode, developers can conduct personalized development based on the business characteristics of the mini - program.

On the same day, Meituan officially announced that it would be the first to access the WeChat AI ecosystem. As part of the first batch of internal testing teams, Meituan had previously jointly developed and tested the access with the WeChat team. In the future, users will be able to use WeChat AI to invoke local life services such as Meituan Takeaway.

Life service platforms such as Ctrip and Tongcheng have also successively announced their access to WeChat.

A few days ago, Tencent Customer Service stated that WeChat is collaborating with mobile phone manufacturers such as Huawei, Xiaomi, Honor, OPPO, and vivo to launch the A2A assistant capability. Currently, several manufacturers have completed the access.

Users can initiate WeChat voice or video calls or send messages to specified friends through the AI assistant of the corresponding mobile phone system.

This is not the first time there has been news about WeChat AI. As early as March this year, foreign media reported that Tencent was promoting a highly confidential AI Agent project within WeChat.

On June 2nd, foreign media reported that Tencent was testing the prototype of the built - in WeChat AI Agent, and the compliance approval process would be launched as early as this month. On the day the report was released, Tencent's stock price closed up 10.5%, with its market value increasing by more than HK$300 billion in a single day, marking the largest single - day increase since January 2021.

WeChat AI may be the ultimate answer for Tencent in the second half of the AI era.

01

The Outline of WeChat AI

People who have seen the early demonstrations revealed that users can swipe right on the WeChat home screen to bring up the AI Agent's chat window. After the user enters an instruction, the Agent will automatically invoke mini - programs in the WeChat ecosystem to complete tasks such as screening, placing orders, and making reservations.

For example, if you say, "Help me order a cup of coffee that costs less than 30 yuan, is not too sweet, and can be picked up nearby," the Agent will automatically invoke mini - programs in WeChat to help you screen coffee shops, match the taste and price, and even complete the ordering process.

Just from the description, it doesn't seem much different from AI chatbots like Doubao and Qianwen.

However, the special thing here is that WeChat AI has the scheduling power over the entire WeChat ecosystem.

Tencent clearly mentioned in its 2025 annual report that its goal is to build next - generation Agentic services within the WeChat ecosystem, connecting the capabilities of mini - programs, content, social interaction, and payment. As of March 31, 2026, the combined monthly active users of WeChat and WeChat International reached 1.432 billion.

In other words, once WeChat AI is launched, whether it is good or bad, it is destined to be a super - app.

There are millions of mini - programs in WeChat, covering daily life scenarios such as ride - hailing, food delivery, ticket booking, and grocery shopping. Almost all of the top domestic Internet services have mini - program entrances in this ecosystem.

The core ability of WeChat AI is to enable AI to invoke the service and transaction capabilities in these mini - programs, completing a complete closed - loop from cognition to decision - making and then to execution.

So how does it do that?

First, understand the user's intention. When a user says, "Help me book a restaurant," the meaning is completely different when said in a family group or a work group.

Who is involved, who can make decisions, how much the budget is, what the dietary restrictions are, and what stage the task is at. These are all the context that the Agent needs to understand. The difficulty lies in the fact that tasks in WeChat naturally span time. A conversation about the summer vacation in a family group may last for several days intermittently.

Then, invoke tools.

The Agent needs to take action, use search engines to find information, use mini - programs to complete queries and price comparisons, use WeChat Pay to complete transactions, and use service notifications to feedback the results to the user.

According to QuestMobile's "2026 Pan - ecosystem Traffic Spring Report", the daily active users of mini - programs have exceeded 900 million, covering hundreds of niche fields.

Now that the toolbox is large enough, the question is whether WeChat AI can use it effectively?

Tencent revealed some technical details in a paper published on March 18th. The WeChat team developed UI - Oceanus, a world model specifically designed for the mini - program ecosystem. Its function is to predict the results of operations. The Agent finds a button, but what will happen when it is clicked? Where will the page jump? What window will pop up? Will the payment process start?

Humans have an intuition about these when operating apps, but the Agent doesn't. So it has to learn from the data.

Game AI learns "how the character will move when this key is pressed," while the world model of mini - programs learns "how the page will change when this button is clicked."

Training directly in the real mini - program environment is too slow and unstable. So UI - Oceanus automatically simulates operations and page changes, generating 5 million samples. This allows the Agent to learn to operate mini - programs in a virtual environment and then transfer to real scenarios.

There is also the cost issue. If the entrance with 1.4 billion monthly active users triggers inference in every scenario, the cost will be astronomical. Tencent needs to balance using small models for basic tasks and invoking powerful models for complex tasks. This multi - model scheduling ability needs to ensure the effect while controlling the cost.

Finally, there is ecosystem coordination.

There are too many mini - programs in WeChat. Service quality, interface stability, merchant cooperation, payment process, recommendation ranking, and profit distribution. Each of these aspects can be discussed at length.

The AI Agent needs to actually complete the tasks for users, not just promise smoothly but then get lost halfway through the operation.

So WeChat AI is actually a very complex project. It has to face various complex scenarios. It needs to understand natural language, invoke mini - programs, handle payments, manage context, and coordinate the ecosystem.

The outline of WeChat AI is clear, but this product will be much larger than we imagined.

02

Why WeChat Is the Most Suitable to Undertake This Agent

The richer the context, the better the AI can understand your real intention, and the more accurate the decisions it makes will be.

And WeChat happens to be Tencent's largest context container.

WeChat has a relationship chain. The social relationships, chat records, and group conversations of 1.4 billion users in WeChat are all context. WeChat has mini - programs, and the service scenarios covered by millions of mini - programs are also context.

WeChat has a payment system. Users' consumption habits, payment records, and transaction preferences are still context.

WeChat has content. The information flows in official accounts, video accounts, and Moments are also context.

Recently, Tencent has launched many AI products, such as Yuanbao, ima, WorkBuddy, and Marvis. They seem to be independent of each other. In fact, they are all accumulating capabilities for WeChat AI.

Behind this is a mechanism called Co - Design within Tencent.

Simply put, Co - Design means that the product team and the model team design and optimize together.

The traditional approach is that the model team first trains the model and then hands it over to the product team for use. When the product team finds problems, it feeds them back, and the model team makes adjustments.

This process is slow, and there is often a situation where "the model is powerful but the product is not user - friendly."

The Co - Design approach is different. The Yuanbao team will tell the Hunyuan team how users actually ask questions and what problems they encounter in real scenarios. The Hunyuan team will specifically optimize certain capabilities of the model based on these real - world feedbacks.

After optimization, the Yuanbao team immediately conducts tests. When new problems are found, they continue to make adjustments.

This process is two - way and synchronous. The product provides real - world data and feedback to the model, and the model provides stronger capabilities to the product.

Why is this approach effective? Because the most fundamental difference between the LLM era and previous AI is generalization.

Before LLM, for a translation product, you only needed to prepare good translation data, and for a Go program, you only needed to prepare Go - related data.

But today, the situation is different. Even if you just want to develop a Coding Agent, you need the model to have chat ability, search ability, instruction - following ability, and reasoning ability. So ultimately, it becomes a very complex interdisciplinary problem.

Tencent's Co - Design with Yuanbao aims to enable the Hunyuan model to have strong chat and search abilities. These abilities can be transferred to other products such as ima and WorkBuddy. The capabilities trained in one product can make other products more user - friendly.

Specifically, Yuanbao deals with the real - world Prompt distribution. The questions users ask in Yuanbao are often vague, maybe just one or two sentences, and they will keep asking follow - up questions.

The multi - round dialogue ability and intention - understanding ability trained in these scenarios can be directly transferred to the context understanding of WeChat AI when handling group chat tasks.

WorkBuddy accumulates data in the office collaboration scenario.

It understands the semantics of enterprise scenarios such as document structure, meeting minutes, and task allocation. These abilities can enable WeChat AI to know how to extract key information and identify decision - making nodes when handling tasks.

ima precipitates search ability. It trains the model on how to convert vague query intentions into precise search strategies and how to filter effective information from a large number of results. These abilities can enable WeChat AI to conduct a round of information screening and intention clarification before invoking mini - programs. In this way, when invoking, WeChat AI will not invoke all possible mini - programs and waste time and tokens, but only invoke a few useful ones.

Marvis trains task decomposition and tool - scheduling abilities.

Marvis decomposes the user's instructions into multiple subtasks and schedules different Agents to operate files, systems, and applications. This task - orchestration and multi - Agent collaboration ability can enable WeChat AI to know how to link mini - program invocations, payment processes, and message notifications when facing cross - scenario tasks such as "Help me order coffee and then notify my colleagues."

These products provide different data, but the data can spread and transfer among each other, forming a network - like system. The data trained in one product can improve the performance of another product through the generalization mechanism of pre - training and post - training.

WeChat AI is now at the center of an AI network.

It doesn't need to start from scratch. It can directly invoke these proven capabilities.

More importantly, WeChat itself is a complete ecosystem. It has a relationship chain, mini - programs, a closed - loop transaction system with WeChat Pay, and a content ecosystem of official accounts and video accounts. These are what other Agent products don't have.

03

How Big Is the Stage for WeChat AI?

All of this is currently achieved through A2A.

A2A stands for Agent - to - Agent, which means "intelligent agent to intelligent agent" in Chinese.

It is an open protocol that stipulates how AI intelligent agents of different manufacturers communicate, invoke capabilities, and ensure security. In contrast, there is the GUIAgent approach, which allows AI to identify the interface by "reading the screen" like a human and then operate WeChat through "simulated clicks."

Tencent's choice of A2A instead of GUI actually has in - depth considerations behind it.

At the Q1 earnings conference call in May, an analyst asked Tencent President Liu Chiping, "How do you view the long - term potential or potential disruption of intelligent agents at the operating system level, including those from iOS, Android, or mobile phone manufacturers?"

Liu Chiping replied, "From the perspective of the operating system, there are several different things mixed here. There are real operating systems like iOS and Android, and then there are applications that try to pretend to be operating systems. If you are an operating system like iOS or Android, you want to ensure that the ecosystem is well - protected and carefully planned, and give applications reasonable permissions. You can have an intelligent agent that tries to provide services for users, but you need to obtain the permission of different applications. Otherwise, as an operating system, you are essentially plundering different applications, which is not the best way to manage an operating system."

What Liu Chiping means is that it is okay for the operating system's Agent to operate applications, but it cannot do so without the authorization of the applications. Otherwise, it is plundering the applications.

In more straightforward terms, Tencent does not accept GUI agents, only A2A.

In the past two years, mobile phone manufacturers have tried to access WeChat externally using GUI.

Honor YOYO promoted the feature of "sending a WeChat red envelope with one sentence," and the selling point of Xiaomi's smart home products was "Xiaomi AI automatically connecting WeChat calls." When you tell your phone, "Send a 10 - yuan red envelope to XX," the AI assistant in the background performs the following operations: unlock the phone, click the WeChat icon, search for XX, click the plus sign, click the red envelope icon, enter 10, and initiate the payment.

This behavior was quickly blocked by WeChat.

In April 2025, the WeChat Security Center issued an announcement, prohibiting third - party tools from bypassing WeChat's security technical measures and illegally obtaining or using WeChat terminal user data.

ByteDance's Doubao phone also suffered the same fate.

In December 2025, the technical preview version of the Doubao phone assistant was released, with the core selling point of "AI directly performing cross - application operations." Soon, a large number of users reported that their WeChat accounts were forcibly logged out, and the system prompted an abnormal login environment. Tencent stated that this triggered WeChat's existing security risk - control strategy.

Honor is the first brand to complete the WeChat A2A adaptation for WeChat AI. Currently, some Honor models support this function. Users can wake up YOYO and directly issue voice commands, such as sending WeChat messages, making WeChat voice or video calls.

A Tencent insider commented that any mobile phone agent that cannot invoke WeChat cannot be considered a real system - level Agent. Tencent will definitely open this door