Meta Muse and WeChat Xiaowei Take Opposite Paths to Personal AI

Meta’s Muse and Tencent’s Xiaowei share a goal: letting an AI take action, not just answer questions. But they start from opposite ends. Muse is presented as a cloud-based assistant users give tasks to across connected services. Xiaowei is built into WeChat, where assistance can appear alongside conversations, content and mini programs. 518

That difference shapes how people encounter each agent, what services it can reach and what kinds of trust it needs to earn. Xiaowei’s availability was still limited in the reporting cited here, so its potential should not be confused with a proven, broad rollout. 811

Muse starts with the task; Xiaowei starts with the app

Muse is designed around delegation: a user gives it a goal, and it can work through a browser and connected services. Meta says Muse runs on a dedicated cloud computer and uses a separate system called Sentinel to review actions that reach the internet; Meta also says the agent asks for permission when needed, including before sensitive actions such as sending an email or making a purchase. Those are the company’s stated safeguards, not a guarantee that an agent will never make a mistake. 1827

Xiaowei takes an embedded approach. It can be called from WeChat contexts such as chats and content, and can connect assistance to mini programs. Reports describe uses including summarizing chat content and helping draft replies. 5 WeChat has also been testing Xiaowei with a small number of users, with text and voice as ways to interact. 811

The practical distinction is where a user begins. Muse asks someone to bring a task to an assistant; Xiaowei can be available within an app they are already using. That may make Xiaowei feel more convenient for WeChat-based activities, while Muse’s design is aimed at tasks that cross services. The available reports do not establish which approach users will prefer over time.

Social coordination brings a different kind of permission

A cross-service assistant can help an individual carry out a task, but social coordination introduces another person’s preferences and consent. Xiaowei’s reported “AI Social” test explores that directly: one assistant can explain a proposed contact to another person’s assistant and request the recipient’s authorization before the two assistants communicate. The feature was described as a small-scale test, not evidence of broadly available or proven social coordination. 2

That design points to a key trust question for assistants in social apps: when may an agent use conversational context, contact someone or speak on a person’s behalf? In the reported test, consent from the recipient is part of the process. 2 For any agent that acts across services, users likewise need to understand what it can access and when it will ask before acting; Meta says Muse requests permission for sensitive actions. 27

Service access shapes what each agent can do

Muse’s cross-service model depends on connections to services beyond its own interface. Its cloud computer and browser give it a way to work across websites, while connected services provide additional access. 18 This makes permissions and provider access central: what the agent can do depends on the connections available and the authority the user grants it.

Xiaowei’s path runs through WeChat’s mini programs. WeChat’s developer guidance describes a process for developers to make agent capabilities available through mini programs, including applying and submitting functions for review. 10 That route can bring services into a familiar environment, but it also means the assistant’s usefulness depends on which providers enable capabilities there.

The commercial implications are different—not settled

An assistant that compares options across outside services could influence which provider a user chooses. That is a possible consequence of Muse’s cross-service design, not an established outcome. 18 For Xiaowei, actions routed through WeChat mini programs make the platform and participating services part of the experience; developer access is subject to WeChat’s process. 10

In either case, providers may care about how an assistant presents choices and directs users. The available sources describe the systems and their access paths, but do not establish how either product will affect competition, merchant relationships or sales.

Trust is the shared test

Muse’s promise depends on users being comfortable giving an agent access to accounts and letting it take consequential actions. Meta says Sentinel reviews outbound actions and that Muse seeks permission for sensitive steps. Users will still need clear boundaries and a way to understand what the agent has done. 27

Xiaowei’s embedded setting raises a different emphasis: the boundaries around conversations, social context and contacting other people. The reported social test’s authorization step is one example of a consent mechanism, but its limited testing means the broader behavior and reach remain uncertain. 28

The contrast is not simply between a standalone agent and an in-app feature. It is between two routes to usefulness: Muse asks users to delegate work across services, while Xiaowei aims to make assistance available inside WeChat’s existing contexts. Neither model removes the hard questions about access, provider cooperation and user control.