WeChat has quietly taken the first step into AI social networking.

Recently, WeChat has launched a small-scale test of a new "Xiaowei AI Social" feature.

If a user wants to contact a friend, they can first tell their own "Xiaowei" about the request. After Xiaowei finds the other user's Xiaowei, it will state the topic of the communication and request authorization from the recipient. Once consent is obtained, the two Xiaoweis will communicate first; when there is a need for the user to make a choice or confirmation, they will respectively go back to remind their respective owners.

This is not the same as "help me send a WeChat message to someone".

When sending a message on behalf of the user, the user still decides what to say, and the AI only clicks send on their behalf. In Xiaowei AI Social, the user gives the AI a goal. The process of initiating inquiries, exchanging information, and sorting out differences is partially completed by the two AIs.

According to actual tests by GeekPark, the exchanges between the two Xiaoweis will not appear in the original WeChat chat boxes of both parties, but remain in an independent space inside Xiaowei. The AI does not directly enter the private conversation, but establishes an agent channel between the two users.

However, Xiaowei is still in the small-scale testing phase at present, and related functions have not been fully opened. Judging from the product maturity, this is still an early experiment. But the problem it touches on is that when AI can represent one person to communicate with another person's AI, the objects connected by social platforms will begin to expand from "person to person" to "person's agent to person's agent".

For WeChat, this is the most advantageous and most cautious AI route that cannot be rushed.

01

The acquaintance relationship chain,

is the most realistic entry point for WeChat to develop Agent

In the past, when AI entered social products, it mainly played the role of content assistance: generating copywriting, polishing replies, or summarizing chat records. AI can help users speak, but rarely truly participates in communication between people.

Since the beginning of this year, product experiments on personal Agents and mutual communication between Agents have been increasing.

OpenClaw has shown more users that AI can connect tools, run continuously and perform tasks on behalf of users; products such as Elys and Second Me try to train digital avatars with users' experiences, interests and expression habits, so that AI can help filter information, find targets and complete initial interactions.

However, to let AI represent users in social interactions, the first problem to solve is where the identity and relationship come from.

For a new platform, users not only need to re-establish relationships, but also provide enough personal information, and trust that AI can accurately express their ideas. Even if two Agents generate a large number of interactions, there is no guarantee that these interactions will eventually become relationships between two real people. AI avatars can be generated quickly, but real trust is difficult to build from scratch.

WeChat does not need to solve this layer of cold start.

As of the second quarter of 2026, the combined monthly active accounts of WeChat and WeChat have reached 1.439 billion. Users' family members, friends, colleagues and customers already exist on the platform. Xiaowei does not need to find a new group of people for users, but can add an agent communication channel based on the existing friend relationships.

This constitutes the most realistic advantage of WeChat in developing AI social. In terms of question answering, search and content generation, Xiaowei faces a large number of general AI products with similar capabilities; what is really difficult for WeChat to replicate is the real identity and acquaintance relationships accumulated over more than ten years. Only by letting AI enter these relationships and services can Xiaowei form different values from independent AI assistants.

Image source: GeekPark

In WeChat, relationships already exist, and what AI needs to do is to help both parties in the relationship reduce the cost of transactional communication. This also makes scenarios such as arranging meals, organizing activities, and confirming itineraries easier to implement. They only require AI to collect information, sort out differences in existing relationships, and return optional solutions to people.

Taking the simple scenario of arranging a meal as an example, two Xiaoweis can now confirm the time and needs for users first, and if they can call restaurant, map, taxi-hailing and payment services in the future, the operations that originally required users to complete between the chat box, search results and multiple mini-programs or platforms may be reorganized into a continuous task.

Martin Lau, President of Tencent, mentioned on the earnings call in August this year that in the future, users can directly send complex instructions to Agents in WeChat; in the longer term, every user, mini-program and merchant may have their own Agent and complete tasks through mutual communication.

"Xiaowei AI Social" may be regarded as an early test of this concept. If it can be realized, WeChat will no longer only be a super App that accommodates chat, payment and mini-programs, but may also become an operating environment for personal Agents to find contacts, coordinate conditions and call services.

This is the most natural direction for WeChat to develop AI.

But the relationship chain is not a data resource that can be arbitrarily scheduled. The reason why users keep their real friends and family on WeChat is that it is first and foremost a system for direct communication between people. The deeper Xiaowei integrates into this system, the more it needs to prove that it will not consume the acquaintance trust that WeChat relies on to establish in order to demonstrate AI capabilities.

02

From Xiaowei to A2A,

WeChat is testing the boundaries of Agent

What is more noteworthy about this "Xiaowei AI Social" test is that it is the first time WeChat has tried to put Agent into the communication link between people.

In the past, AI in WeChat was more responsible for tasks such as searching, summarizing, generating copywriting or performing operations. Even for "help me send a message", the main body of communication is still the user, and AI only generates or delivers the content on their behalf.

This test takes a step forward: users only need to express their goals, and their own Xiaowei can contact the friend's Xiaowei; after getting the other party's authorization, the two Agents will exchange information first, and when it comes to links that require commitment and choice, they will invite people back to confirm.

The communication link thus becomes "Human — Agent — Agent — Human". AI is no longer just the interaction entrance between people and software, but also begins to become an intermediary between two users.

Looking at it alone, this is still only a feature in small-scale testing, and its applicable scenarios are quite limited. But if we put it back into the update context of Xiaowei since the beginning of this year, the outline of WeChat's Agent product is much clearer than it was a few months ago.

In June this year, WeChat began to expand the gray-scale test of Xiaowei. It is placed in the upper left corner of the WeChat homepage, users can communicate with it through text or voice, and also let it search for materials, summarize files, process images, set reminders, or call mini-programs to complete tasks such as ordering food, hailing a car, and booking tickets. At the same time, WeChat began to open up AI ecosystem access capabilities to mini-program developers, so that Xiaowei can not only answer questions, but also find and call services inside WeChat.

After that, AI was successively embedded in specific scenarios such as Official Accounts, Moments, chat images and Channels. It can summarize articles, assist in processing images, and also try to generate Moments copy and comments. Up to now, testing A2A communication between Xiaoweis, WeChat's attempts at AI have extended from information acquisition to content generation, from calling services to cross-user coordination.

This product context is becoming clear: what WeChat wants to do may be an AI interaction layer that can understand user intentions, call WeChat's capabilities, and complete tasks between different services and users. In the past, mini-programs put a large number of services into WeChat; Xiaowei may try to let users no longer find and operate these services one by one, but directly put forward goals through natural language.

However, Xiaowei is still in the gray-scale testing phase at present, many functions are scattered in different entrances, and the performance of capabilities is not stable. For the vast majority of ordinary users, it has not yet become a clearly perceptible way of use; sometimes, the AI entrance even interferes with the original operations. For example, when searching for relevant information after watching content on Channels, the original clear search demand may be directed to "Ask AI". When AI does not give more accurate and efficient results, this change is more like the product is looking for usage scenarios, rather than users really needing an Agent.

This may also be a contradiction in WeChat's current AI attempts: on the one hand, it needs to verify as soon as possible whether Agent can enter a real, high-frequency usage environment; on the other hand, WeChat already carries communication, content, payment and life services, and any new entrance may change the operation path familiar to users. If AI is embedded too shallowly, it is difficult to form independent value; if embedded too deeply, it may disturb the original experience, and even touch the boundary of social relationships.

The A2A function pushes this contradiction to a more sensitive position.

Image source: GeekPark

Judging from the existing functions, WeChat has actually set up two different external communication paths for Xiaowei.

The first is to send messages on behalf of users. Users can let Xiaowei send WeChat messages directly to friends, which are confirmed by the user before being sent, and then the messages enter the original chat window of both parties. In this process, Xiaowei mainly performs the execution function: the expression content comes from the user, the AI is responsible for completing the sending, and the other party does not need additional authorization.

The newly tested A2A function is different. The user does not let Xiaowei send a message to the friend himself, but lets his own Xiaowei contact the friend's Xiaowei. After the initiator confirms, the other party also needs to decide whether to allow their own Xiaowei to participate, and the two Agents can continue to communicate in an independent space; when there is a link that requires a choice, Xiaowei will confirm with the users respectively.

Image source: GeekPark

What is distinguished here may be the boundary between "AI executes human instructions" and "AI represents people to participate in negotiation".

This caution is related to the environment WeChat is in. WeChat has an acquaintance relationship chain that is difficult for other AI products to replicate, so it is closer to real interpersonal relationships than general tools. If the Agent misunderstands an affair, the user can re-operate; but if it misjudges the tone, position or weight of a sentence, it may not only lead to task failure, but also cause misunderstandings between people.

This also means that Xiaowei still needs to answer a more critical question than "what other functions can be added" next: who on earth should AI be in WeChat?

It can be a relatively independent personal assistant to help users search for information and complete tasks; it can also become an AI operation assistant for WeChat accounts to assist in handling messages, content and contacts; it can also become a LUI entrance that uniformly schedules WeChat services, allowing natural language to gradually replace menus and buttons. Taking a step further in the future, it may even be shaped into a digital avatar with long-term memory and personality.

These different positions correspond to completely different permissions, identities and responsibility boundaries. Especially in A2A negotiation, Xiaowei needs to understand the user's demands and preferences. How long this information will be stored, how many scenarios it can be reused in, and to what extent it will be disclosed to the other Agent will all affect whether users are willing to hand over real affairs to it. The ability to complete communication technically does not mean that it has gained the trust to communicate on behalf of users.

Therefore, the A2A function is more like a boundary test of WeChat for the role of Agent. For WeChat, the real long-term challenge may be to draw a clear line between what things can be handed over to it and what things must be done in person.

This article is from the WeChat Official Account "GeekPark" (ID: geekpark), Author: Lian Ran, Editor: Zheng Xuan, published with authorization from 36Kr.