WeChat AI Agent Launches, Major Internet Platforms Respond

The WeChat agent is really coming.

The WeChat Open Platform has released the following content regarding the WeChat AI developer guidelines.

Guidelines state that, to provide users with a more intelligent interactive experience and help them more efficiently discover and use mini-program services, the WeChat Open Platform offers developers convenient access to the WeChat AI ecosystem, while fully respecting developers' rights and autonomous choices.

After integration, the mini-program will have the opportunity to be recommended and invoked by WeChat AI. Mini-programs that have not been integrated cannot be invoked by WeChat AI.

The platform offers two integration modes. In automatic mode, the platform is authorized to access the mini-program source code during review, requiring no additional development. In development mode, developers can customize and personalize the mini-program based on its business features.

On the same day, Meituan officially announced it was the first to integrate with WeChat's AI ecosystem. As one of the first beta testing teams, Meituan had previously collaborated with the WeChat team to jointly develop and test the integration. In the future, users will be able to access Meituan's local life services, such as food delivery, through WeChat AI.

Travel platforms such as Ctrip and Tongcheng have also successively announced their integration with WeChat.

Several days ago, Tencent customer service stated that WeChat is collaborating with smartphone manufacturers such as Huawei, Xiaomi, Honor, OPPO, and vivo to launch A2A assistant capabilities, and several manufacturers have already completed integration.

Users can initiate WeChat audio or video calls or send messages to designated contacts using the AI assistant of their respective mobile operating system.

This is not the first time there have been rumors about WeChat’s AI. Back in March this year, foreign media reported that Tencent was advancing a highly confidential AI Agent project within WeChat.

On June 2, foreign media reported that Tencent is testing a prototype of an AI agent integrated into WeChat, with compliance approval procedures potentially beginning as early as this month. On the day the report was published, Tencent's stock closed up 10.5%, adding over HK$300 billion to its market capitalization in a single day—the largest single-day gain since January 2021.

WeChat AI may be the ultimate answer for Tencent AI's second half.

People who have seen the early demo revealed that users can swipe right on the main WeChat interface to open the AI Agent chat window. After entering a command, the Agent automatically triggers mini-programs within the WeChat ecosystem to complete tasks such as filtering, placing orders, and making reservations.

For example, if you say, “Help me order a coffee under 30 yuan, not too sweet, and available for pickup nearby,” the Agent will automatically access the mini-program within WeChat to filter coffee shops, match your taste and price preferences, and even complete the order for you.

Based on the description alone, it’s no different from AI chatbots like Doubao or Qwen.

However, the unique aspect here is that WeChat AI has control over the entire WeChat ecosystem.

In its 2025 annual report, Tencent explicitly stated its goal to build the next generation of agentic services within the WeChat ecosystem, integrating mini-programs, content, social, and payment capabilities. As of March 31, 2026, the combined monthly active users of WeChat and WeChat Outside China reached 1.432 billion.

In other words, once WeChat AI launches, it is destined to become a super app—whether for better or worse.

WeChat has millions of mini-programs covering everyday scenarios such as ride-hailing, food delivery, ticket booking, and grocery shopping. Nearly all leading domestic internet services have a mini-program entry point within this ecosystem.

The core capability of WeChat AI is to enable AI to access services and transaction functions within these mini-programs, completing a full cycle from cognition to decision-making to execution.

So how does it work?

First, understand the user’s intent. When a user says, “Help me book a restaurant,” the meaning differs entirely when said in a family group versus a work group.

Who is involved, who has decision-making authority, what the budget is, what the constraints are, and how far along the task has progressed—these are all contexts the Agent needs to understand. The challenge lies in the fact that tasks on WeChat naturally span time; conversations in a family group about summer vacation may stretch intermittently over several days.

Then, call the tool.

The agent needs to take action: search for information, complete inquiries and price comparisons using mini-programs, process payments via WeChat Pay, and deliver results to the user through service notifications.

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

The toolbox is large enough now; the question is, does WeChat AI understand how to use it?

In a paper released on March 18, Tencent revealed some technical details. The WeChat team developed UI-Oceanus, a world model specifically designed for the mini-program ecosystem. Its purpose is to predict the outcomes of actions. The agent identifies a button—but what happens when it’s clicked? Which page will it navigate to? What pop-up window will appear? Will the payment process be triggered?

Humans have an intuitive understanding of these when using the app, but agents lack this intuition, so they must learn from data.

The game AI learns "pressing this button makes the character move this way," while the mini-program's world model learns "clicking this button changes the page this way."

Training directly in the real mini-program environment is too slow and unstable, so UI-Oceanus automatically simulates user interactions and page changes to generate 5 million samples. This allows the Agent to learn how to operate mini-programs in a virtual environment before transitioning to real-world scenarios.

There’s also the issue of cost. If inference is triggered in every scenario for an entry point with 1.4 billion monthly active users, the cost would be astronomical. Tencent must strike a balance between using smaller models for basic tasks and invoking stronger models for complex tasks. This multi-model orchestration capability must ensure performance while controlling costs.

Finally, ecosystem coordination.

There are too many mini-programs in WeChat; service quality, interface stability, merchant cooperation, payment processes, recommendation rankings, and profit distribution—each of these alone could be discussed at length.

An AI agent must actually get things done for users—it can't just sound smooth in its promises and then get lost halfway through placing an order.

Therefore, WeChat AI is actually a highly complex engineering system that must handle a wide range of intricate scenarios. It needs to understand natural language, invoke mini-programs, process payments, manage context, and coordinate across the ecosystem.

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

The richer the context, the better the AI can understand your true intent and make more accurate decisions.

WeChat is precisely Tencent's largest contextual container.

WeChat has a social network, with 1.4 billion users whose social connections, chat histories, and group conversations all constitute context. WeChat also has mini-programs, with millions of mini-programs covering diverse service scenarios, which also constitute context.

WeChat has payment capabilities, and user spending habits, payment records, and transaction preferences are still contextual.

WeChat contains content, including information streams from official accounts, video accounts, and Moments, all of which are context.

A while ago, Tencent launched many AI products, such as Yuanbao, Ima, WorkBuddy, and Marvis, which appear to be independent. In reality, they are all building capabilities for WeChat AI.

Behind this is Tencent’s internal Co-Design mechanism.

In simple terms, Co-Design means the product team and the model team design and optimize together.

Traditionally, the model team would first train the model and then hand it off to the product team for use. If the product team encountered issues, they would provide feedback, and the model team would make adjustments accordingly.

This process is slow, and it often results in "a strong model but a poorly usable product."

The Co-Design approach is different. The Yuanbao team informs the Hunyuan team about how users actually ask questions in real-world scenarios and what issues they encounter. Based on this real-world feedback, the Hunyuan team specifically optimizes certain capabilities of the model.

Optimization is complete. The Yuanbao team will immediately proceed with testing and make further adjustments if new issues are discovered.

This process is bidirectional and synchronized. The product provides the model with real data and feedback, while the model enhances the product’s capabilities.

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

Before LLMs, creating a translation product only required high-quality translation data, and developing a Go program only required well-prepared Go data.

But today is different—even if you just want to build a coding agent, you need the model to have conversational ability, search capability, instruction-following skills, and reasoning ability. As a result, it ultimately becomes a highly complex interdisciplinary problem.

The co-design between Tencent and Yuanbao aims to give the Hunyuan model strong chat and search capabilities, which can then be transferred to other products such as Ima and WorkBuddy. Skills trained in one product can enhance the usability of other products.

Specifically, Yuanbao handles real-world prompt distributions. Questions users ask in Yuanbao tend to be vague, often just one or two sentences, and users frequently follow up with additional questions.

The multi-turn dialogue and intent understanding capabilities trained on these scenarios can be directly applied to context comprehension when handling group chat tasks with WeChat AI.

WorkBuddy accumulates data from office collaboration scenarios.

It understands the semantics of enterprise scenarios such as document structure, meeting minutes, and task assignments. These capabilities enable WeChat AI to know how to extract key information and identify decision points when handling tasks.

IMA encapsulates search capabilities. It trains models to transform ambiguous query intentions into precise search strategies and to filter out relevant information from vast volumes of results. These capabilities enable WeChat AI to perform an initial round of information filtering and intent clarification before invoking mini-programs, ensuring that it does not unnecessarily call all potentially relevant mini-programs—thereby saving time and tokens—and instead invokes only a select few that are truly useful.

Marvis is trained in task decomposition and tool scheduling.

Marvis breaks down user instructions into multiple subtasks and coordinates different agents to control files, systems, and applications. This task orchestration and multi-agent collaboration capability enables WeChat AI to handle cross-context tasks like “Order coffee for me and notify my colleagues” by seamlessly connecting mini-program invocations, payment processes, and message notifications.

These products provide different types of data, but the data can be shared and transferred among them, forming a network-like system. Data trained on one product can enhance the performance of another through pre-training and post-training generalization mechanisms.

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

It doesn't need to start from scratch—it can directly leverage these already validated capabilities.

More importantly, WeChat itself is a complete ecosystem, featuring social connections, mini-programs, a closed-loop transaction system with WeChat Pay, and a content ecosystem through official accounts and video accounts—features that other agent products lack.

All of this is currently powered by A2A.

A2A stands for Agent-to-Agent, known in Chinese as "intelligent agent to intelligent agent."

It is an open protocol that defines how AI agents from different manufacturers communicate, invoke capabilities, and ensure security. In contrast, the GUIAgent approach enables AI to interact with WeChat by “reading the screen” to recognize interfaces and “simulating clicks” to perform actions, just as a human would.

Tencent chose A2A over GUI, and this decision was based on careful consideration.

During Tencent's Q1 earnings call in May, an analyst asked Tencent President Liu Chiping, "How do you view the long-term potential or potential disruption from agents at the operating system level, including those from iOS, Android, or smartphone manufacturers?"

Liu Chiping replied, “From an operating system perspective, this mixes several different things. There are true operating systems like iOS and Android, and then there are applications that try to pretend they are operating systems. If you are an operating system like iOS or Android, you want to ensure the ecosystem is well-protected and carefully curated, granting applications reasonable permissions—you can have agents that serve users, but you need permission from different applications. Otherwise, as an operating system, you’re essentially exploiting different applications, which is not the best way to manage an operating system.”

Liu Chiping meant that it is acceptable to use an operating system agent to control applications, but only with the application's authorization; otherwise, it amounts to plundering the application.

In simpler terms, Tencent doesn’t accept GUI agents—only A2A.

Over the past two years, smartphone manufacturers attempted to integrate WeChat externally via GUI.

Honor YOYO advertised "send WeChat red envelopes with a single command," and Xiaomi's smart home products highlight "Xiao Ai automatically connects WeChat calls." When you tell your phone, "Send a 10-yuan red envelope to XX," the AI assistant performs the following actions in the background: unlock, tap the WeChat icon, search for XX, tap the plus sign, tap red envelope, enter 10, and initiate payment.

This behavior was quickly banned by WeChat.

In April 2025, WeChat Security Center announced a ban on third-party tools that circumvent WeChat's security measures to illegally obtain or utilize data from WeChat end users.

The ByteDance DouBao phone has met the same fate.

In December 2025, the technical preview of DouBao Mobile Assistant was released, with its key feature being "AI-driven cross-app operations." Soon after, numerous users reported being forcibly logged out of their WeChat accounts, with system notifications indicating abnormal login environments. Tencent stated that this triggered WeChat’s existing security risk control policies.

On WeChat AI, Honor is the first brand to complete WeChat A2A integration. Currently, select Honor devices support this feature, allowing users to activate YOYO and issue voice commands, such as sending WeChat messages, making voice calls, or initiating video calls via voice.

A Tencent insider remarked that any mobile agent that cannot access WeChat cannot be considered a true system-level agent. Tencent will definitely open this door—it’s just a matter of time.

WeChat is willing to allow smartphone manufacturers' agents limited access to WeChat capabilities through controlled protocols such as A2A, but will not permit external agents to enter WeChat via screen reading or simulated clicks.

This shows that, ultimately, Tencent still needs to control the access rights and rule-making authority for the WeChat ecosystem.

When it comes to DouBao, this raises another question: Will WeChat AI charge fees?

Doubao has 345 million monthly active users, and recent reports suggest it will soon start charging for certain features; WeChat, with 1.4 billion monthly active users, faces even greater pressure.

Moreover, with so many users, if the WeChat AI triggers inference for every scenario, the cost would be astronomical.

The previously rumored 10 billion yuan investment by Tencent in DeepSeek can be understood as a foundation for model supply and cost infrastructure.

Tencent's self-developed HunYuan large model needs technical partners, and the WeChat ecosystem particularly requires low-cost inference capabilities. DeepSeek's low-cost training approach aligns perfectly with the needs of WeChat AI in scenarios involving massive user volumes.

On June 2, Tencent also announced that the invocation prices for the DeepSeek-V4 series on the Tencent Cloud platform are fully aligned with DeepSeek's official pricing, with users bearing no cloud platform premium.

All these clues suggest that Tencent is deeply aligning with DeepSeek, and the WeChat Agent is likely the first outcome of this partnership.

Use small models for basic tasks to keep costs low and speed high. For complex tasks, invoke powerful models to achieve better results and higher accuracy. This multi-model scheduling capability ensures optimal performance while controlling costs.

As a WeChat user, if WeChat AI could truly complete a task correctly the first time, I would be willing to pay for this capability.

For example, booking a flight for me, finding a restaurant for me, and reminding me who that person is—someone who’s been sitting in my social media list for ages with whom I’ve never had a conversation. I think these features are all very valuable.

More importantly, WeChat AI serves not only individual users but also enterprise customers. Scenarios such as enterprise automation, intelligent customer service, and intelligent marketing have stronger demands for AI and higher willingness to pay.

The stage for WeChat AI is actually very large. Just how large? The answer is: the stage for WeChat AI is as big as the WeChat ecosystem.

Yao Shunyu, Chief AI Scientist at Tencent and lead of the Hunyuan large model, provided a more long-term perspective at the Tencent Cloud AI Industry Application Conference on June 5.

He believes that AI is a long-term game, not a short-term window. He criticized the mindset of some Silicon Valley practitioners who aim to "make money fast and retire in two years," emphasizing that we are currently in a phase similar to "the 1970s personal computer era," where new product opportunities will continue to emerge.

This insight precisely explains why Tencent is willing to invest so heavily in WeChat AI. Yao Shunyu emphasized, “Practical value outweighs leaderboard value.” He believes that AI methodologies are already highly mature; the real challenge lies in identifying the right problems to solve, rather than chasing numbers on rankings.

WeChat AI is precisely addressing such a "good question."

How can we enable these 1.4 billion users to experience the value of AI in their daily lives?

There’s no flashy gimmicks or leaderboard chasing—only by solving this important problem can Tencent truly enter the second half of AI.

This article is from the WeChat public account "Face AI" (ID: faceaibang), authored by Miao Zheng and edited by Wang Jing.