WeChat AI has no intention of taking over everything
If one were to pick an "AI straggler" among the major domestic tech companies, many would choose Tencent.
The popularity and influence of its models are far inferior to those of Alibaba and ByteDance. The user base of Yuanbao lags behind that of Doubao, and the industry reputation of Hunyuan lags significantly behind that of Qianwen.
It wasn't until June 2026 that WeChat, Tencent's biggest trump card, took substantial action.
Starting from June 8th, Tencent officially opened up the ability to access its AI ecosystem and released access guidelines for the "mini-program AI development mode" to developers.
The first batch of internal testing list includes several leading platforms. Currently announced applications for access include JD.com, Meituan, Ctrip, Didi, Tongcheng, Dewu App, etc., covering core scenarios such as e-commerce, food delivery, travel, and tourism.
The long-awaited WeChat AI finally took shape.
However, from the perspective of the ecological access arrangement and calling logic of WeChat AI, there is a key choice that seems ordinary but has been overlooked:
WeChat did not choose to let AI fully take over mini-programs. Instead, beyond automatic operations, it allows developers to define what AI can access on their own.
According to the access guidelines, developers need to actively complete the access application, ability declaration, skill encapsulation, review submission, and opening of functions related to mini-programs.
WeChat's Agent is only responsible for understanding user intentions, selecting capabilities, scheduling execution, and returning results. The official documentation positions it as "fully respecting the rights and autonomous choices of developers."
At a time when everyone is competing to see whose Agent is stronger and more capable of replacing everything, WeChat chose a path that is not that glamorous.
This choice itself hides WeChat and even Tencent's confirmation of the relationship between AI and the application ecosystem.
The Path to Applications
In the past year, Agent has been the hottest thing in the entire AI industry. Almost all large companies are talking about the same thing: AI should help users complete tasks.
As Agent capabilities become stronger and stronger, and even in the context of the panic of AI killing software, a contradiction between the new AI world and the old Internet world has emerged: Where does the application side stand?
Currently, the industry generally presents two approaches.
One can be called AI actively infiltrating applications, that is, understanding applications, operating applications, and becoming applications.
Whether it's OpenAI's Operator, Anthropic's ComputerUse, or the mobile phone Agents launched by some domestic automakers, the core technical route is the GUI (Graphical User Interface) Agent combined with the VLM model, which allows AI to read the screen, understand the interface, and simulate clicks. When the SDK or API cannot be obtained, it can bypass the cooperation and authorization of the application side.
In theory, as long as AI is strong enough, it can bypass all intermediate layers and directly complete operations for users.
The other approach is the opposite, which is to promote applications to actively adapt to AI, but still leave the choice to the application side itself.
Representatives include Apple's AppIntents, Amazon's AlexaSkills, and WeChat's current solution.
The common feature is that developers actively declare capabilities, and AI is only responsible for scheduling the declared capabilities. The application side retains full control, and developers can decide what to open and what not to open.
These two approaches actually answer the question of benefit distribution brought about by technological changes:
To what extent does Agent need to respect the authorization, brand, and commercial interests of the application side while helping users complete tasks?
This question has already caused real and intense frictions in the past six months.
In December 2025, the Doubao mobile phone was launched. This device equipped with an AI mobile assistant can automatically control third-party applications on the phone through the GUI.
However, just three days after its launch, mainstream apps such as WeChat, Alipay, and Meituan imposed varying degrees of restrictions and risk control blockades on it. Some bank clients also popped up security prompts or simply crashed actively.
Looking back, this may be a wall that domestic Agents hit when trying to take over applications.
The China Academy of Information and Communications Technology led the release of the "Security Guidelines for Dual Authorization of End-Cloud Collaborative Intelligent Agent Interaction," clearly proposing a key principle: Agents need to obtain both application authorization and user authorization to legally access third-party applications.
Whether AI can control applications depends not only on whether the model is strong enough but also on whether the application side is willing to be controlled.
In high-trust scenarios such as finance, payment, and healthcare, the application side has sufficient reasons and rights to reject unauthorized automated operations.
WeChat's current solution itself is responding to this question, and the entire access design can be summarized into several things.
First, it gives mini-program developers the right to choose whether to access AI.
Accessing the WeChat AI ecosystem is an active application behavior and is not mandatory by the platform. Developers can choose not to access and continue to maintain the existing form of mini-programs.
Second, the ability boundary is defined by developers themselves.
Developers can decide which Skills to encapsulate and which interfaces to expose on their own. WeChat AI cannot call capabilities beyond the developer's authorization.
Third, the design of the development mode retains the brand presence of mini-programs.
After the encapsulated capabilities are called, the results are rendered into GUI cards through atomic components and embedded in the conversation flow for display.
This means that the brand and interaction mode of mini-programs are still retained in the AI calling link. Users can still see the visual interface of mini-programs, rather than just a text answer without a brand logo.
Fourth, WeChat still retains the conditions for walking on two legs, simultaneously opening the automatic mode and the development mode.
The automatic mode allows WeChat to read the mini-program source code during the review process, analyze the pages, and allows WeChat AI to directly operate the pages. The development mode is for developers to actively declare capabilities. Both modes can be enabled simultaneously.
The final opening indicates that Tencent has no conclusion on the ultimate outcome of Agent internally. It retains the platform-led automated route and also builds a route for developers to actively adapt.
Actually, before the opening of this set of AI capabilities, WeChat had already started to cultivate the supply side. In January 2026, WeChat launched the growth plan for AI applications and online tool mini-programs, covering the whole year. It provides developers with a six - month free cloud development environment, a quota of 100 million Tokens for the Hunyuan model, a quota of 10,000 text - to - image generations, as well as traffic support and commercialization support.
Subsidizing the supply side first, getting developers ready first, and finally opening the entrance is an expression of WeChat returning the AI choice to the application ecosystem.
Values or Pragmatism
WeChat's business choice has nothing to do with morality because its situation is fundamentally different from that of other AI companies.
The blueprint for WeChat's Agent was first revealed by Liu Chiping, President of Tencent, in a conference call in the third quarter of 2025. He said at that time that within WeChat's communication, social, content, mini-program, and even payment systems, an Agent function is expected to become an ideal assistant for understanding user needs and executing tasks.
Tencent obviously hopes to complete everything from understanding needs, matching services, completing transactions, to payment settlement within the WeChat ecosystem.
Tencent's difference lies in its huge ecosystem.
As of the end of March 2026, the daily active users of WeChat mini-programs exceeded 610 million, and the monthly active users reached 973 million. There were more than 7 million registered mini-programs in total, covering almost all service categories such as catering, retail, travel, healthcare, and government affairs. Behind these are millions of developers, merchants, and service providers, which constitute the core value of WeChat as a super app. However, most of these services do not belong to Tencent itself.
If WeChat AI is strong enough to directly help users complete tasks such as booking air tickets, ordering takeaways, and making medical appointments, the interaction mode between users and mini-programs will be fundamentally changed, and the brand, pages, membership system, and developer value of mini-programs will also be reconstituted.
The stronger the Agent capabilities, the greater the impact on WeChat's existing ecosystem.
OpenAI, Anthropic, etc. don't need to consider this problem. Model - first doctrine often prioritizes user goals, that is, whatever the user wants, the Agent capabilities need to be improved to cover it. The software side as the old world doesn't need to be taken care of.
However, as the platform side, WeChat must balance a triangular relationship of "users need to complete tasks efficiently, developers need to retain their brand and commercial value, and the platform needs to maintain the healthy ecosystem."
If WeChat AI directly bypasses developers and takes over all service entrances, the biggest victim may be the WeChat ecosystem itself.
What Zhang Xiaolong said at the WeChat Open Class in 2019 is still valid when re - understood in the context of Agent today:
"The mission of mini-programs is to let creators realize their value and get rewards. Just because we have traffic doesn't mean we should distribute it."
"If we don't decentralize, Tencent will monopolize the top - tier mini-programs, and there will be no room for external developers. It may seem that Tencent can make short - term profits, but the ecosystem will be gone."
These words were the starting point of the mini-program era, but the logic also holds in the Agent era: If WeChat AI becomes the only service scheduling center, can the commitment of "decentralization" still be continued?
Of course, Tencent may not choose to be restrained actively but has to be. It already has a large developer and merchant ecosystem. If WeChat AI directly bypasses developers, it may trigger a backlash from the ecosystem.
Today's "fully respecting the rights of developers" comes from both values and real constraints.
Unique Advantages Hard to Replicate
Behind WeChat's more open approach is an existing, large - scale, and standardized third - party application ecosystem.
First, the existing developer base.
When WeChat invites developers to encapsulate callable capabilities for AI, it is facing a developer network of more than 7 million mini-programs covering almost all categories of life services. The density and breadth of this supply side are difficult for other large companies or AI platforms to reach.
Second, user habits and scenario trust.
Users are used to completing services in WeChat, such as ordering takeaways, making medical appointments, paying utility bills, taking the bus, and buying movie tickets. The AI - enabled execution doesn't cause a scenario shift, only the way of opening changes. On the contrary, if services are called and interactive cards are opened in other ChatBot interfaces, it will obviously cross a higher user trust threshold.
Third, the business closed - loop.
Payment, membership systems, and transaction infrastructure are all wrapped within WeChat. AI can call them without friction and doesn't need to jump to external systems. From intention understanding to execution, delivery, and payment, AI forms a complete closed - loop.
Speaking of the problem of Tencent's lagging model capabilities that the market is most concerned about, a problem masked by the current AI narrative is that for many life and consumption - related applications more closely related to WeChat mini-programs, their Agent - based transformation actually has limited requirements for model capabilities. Even using more advanced models to handle these requirements is an uneconomical approach.
Assume that a user uses WeChat AI to book an air ticket. The biggest bottleneck of the task is often not the model parameters and ranking - brushing ability. Instead, the greater barriers are whether it can connect to the airline's inventory, make effective payments, handle ticket refunds, changes, and cancellations in a timely manner, or remember user information.
This means that compared with simply training and ranking, the operation and co - existence of the application ecosystem are more important.
A not - so - leading model connected to 7 million mini-programs may create more real value than the strongest model that cannot connect to any services.
WeChat actually did another thing: It left the most important question in the Agent era to the market to answer.
In the past year, the default premise in the industry was that AI would eventually take over all applications, but this premise has never been verified on a large scale.
Whether users are willing to let AI book air tickets, choose restaurants, make medical appointments, and do shopping on their behalf may vary greatly in different scenarios.
By allowing developers to independently open up capabilities and users to independently choose to use them, WeChat is essentially building an experimental field in a real business environment.
What ultimately remains may not be what AI can do, but what users are willing to let AI do.
In consumer - level scenarios, after the model capabilities exceed the "sufficient" threshold, the marginal return decreases.
In fact, there are many business cases where value is created with not - so - leading technology in the right ecological niche. For example, Nintendo's GameBoy used the worst screen of its time but still outperformed others. Another example is Li Auto, which used the range - extender technology regarded as backward in the industry but achieved large - scale production and was emulated by its peers. They all achieved commercial verification in their own ecological niches.
Of course, having sufficient model capabilities is still the biggest prerequisite.
If the Hunyuan or other models behind WeChat AI cannot understand complex intentions, handle multi - round conversations, or perform accurate intention routing, even the strongest ecosystem cannot make up for the poor user experience.
It's not easy to achieve this, but in today's highly competitive Agent narrative, the ecological thickness may have greater differentiation.
Different Approaches to Problem - Solving
When facing the question of "how Agent and applications should co - exist," large companies have given different answers.
For example, Alibaba Cloud's Bailian opens up the Agent infrastructure. It includes large models, workflow orchestration, plugin ecosystems, full - lifecycle MCP services, and Agent development frameworks, along with supporting capabilities such as knowledge bases and run - times. The core is to solve the problem of how developers can build, deploy, and operate Agents, which transforms the tool layer.
Ant and Alipay open up payment capabilities, including payment MCP, transactions, refunds, risk control, and split - payment settlement. They also launched the ACT intelligent agent business trust agreement, enabling Agents to complete transactions safely and establish business collaboration standards, which transforms the transaction infrastructure.
In the overseas market, Google and OpenAI are more radical. Google hopes that Gemini will become the Agent layer above Android, and OpenAI hopes that ChatGPT will become the new Internet entrance. Their ideas are closer to "Agent - first," that is, user needs come first, and applications are just execution tools.
On the contrary, WeChat may draw more inspiration from Apple in its AI route design.
Apple's AppIntents is also a framework for developers to actively declare capabilities, and the system is responsible for scheduling. It has a powerful third - party application ecosystem, the App Store, and also doesn't want the AI layer to directly destroy the application distribution system.
Due to the similar ecological niches, both Apple and Tencent face the same problem: How to make Agent improve user efficiency without harming the existing developer ecosystem.
The difference between WeChat and Apple is that Apple's third - party application data is still on the application side's server, and the interface docking involved in AI calls needs to be re - engineered. However, mini - program applications run within the WeChat ecosystem, and the friction in code submission, review, and recognition is smaller. Therefore, the "automatic mode" of WeChat AI developers may have greater potential in the future.
From the ways Apple and WeChat handle the relationship between AI and applications, it can be seen that this idea is not unique to Tencent. All platforms with strong ecosystems will face structural challenges in the Agent era.
For mobile phone manufacturers, the idea of WeChat AI may also have certain reference value.
Currently, many domestic AI mobile phones still rely on the GUI route of OCR, screen understanding, and simulated clicks. However, the practice in the past six months has proven that this route will face risk - control barriers in high - trust scenarios such as banking, payment, and social media.
If a "capability declaration first, GUI as a fallback" model is formed in the future, that is, giving priority to calling the capabilities actively declared by applications and using the GUI as a fallback when the capabilities do not exist, the execution success rate and security may be much higher than the pure GUI route.
Although Google is bold, Android is also moving