Meta Muse’s 2.5 million reported downloads do not settle the AI agent race

How do Muse, dots and Xiaowei reach other services?

An AI agent, in this context, is software designed to take actions through apps or online services, rather than only respond to a prompt. Wednesday’s Semafor report describes Muse and dots as assistants built to work with services beyond their makers’ platforms. Tencent’s Xiaowei is embedded in WeChat and draws on its mini-program ecosystem: services that run inside the app.

Tencent described Xiaowei as a prototype intended to help users discover products, make purchase decisions and place orders through WeChat. The company said the prototype uses user intervention and multiple confirmations, with broader access planned in phases. The supplied reporting gives no user count or task-completion rate for Xiaowei, and does not establish broad public availability.

That difference affects where each agent can operate. Xiaowei can draw on services available in WeChat, but that focus may constrain its reach elsewhere. Semafor reported that other Chinese platforms have blocked outside agents, creating access barriers for systems that need to work across separate apps.

Semafor also cited the newsletter Hello China Tech, which reported mistakes in early Xiaowei tests and questioned whether transactions would bring Tencent new revenue. The account does not establish how representative those tests were or how often Xiaowei can complete purchases successfully.

What do Muse’s downloads and product descriptions show?

The South China Morning Post reported that Muse accumulated 2.5 million downloads in its first two weeks. That is a reported figure, not an official Meta total. It does not show how many people used Muse repeatedly, how many tasks it completed or how it handled errors.

Meta announced Muse on Sept. 8. The company says the agent can work across connected apps, run in the background and request approval for sensitive actions. Those are Meta’s descriptions of its product and safeguards, not findings from an independent evaluation.

OpenAI announced dots on Sept. 29 and said it was beginning a phased rollout for eligible Pro, Business Premium and Enterprise users in selected markets. OpenAI describes dots as agents with cloud computers that can use connected apps. Its product descriptions and examples are not independent evidence of how well dots performs on everyday tasks.

The milestones are not interchangeable. The reported Muse number is a download figure, while Tencent’s account of Xiaowei describes a prototype and OpenAI’s announcement describes a phased rollout. None of those facts, by itself, establishes how reliably an agent completes a task.

What access do cross-app agents need?

Agents that work across outside services depend on access to those services. Semafor reported that Muse and dots require users to provide access to private systems. That is an access requirement, not evidence that either product routinely fails to use connected services.

Semafor also reported that Amazon had blocked Muse, but did not establish the scope of that block or how often it prevents task completion. The report describes access constraints on both sides of the comparison: Chinese platforms have blocked outside agents, while agents designed to work across services also need permission to reach private systems.

For users, connecting an agent to an app is only one step. Access may allow software to reach a service without showing that it can complete a multi-step action accurately or with appropriate oversight. The reporting supplies no comparable completion rates for Muse, dots and Xiaowei.

What evidence would show which agent works better?

The reporting reviewed here provides no independent, like-for-like test of Muse, dots and Xiaowei on the same tasks. It supports a comparison of product strategies and reported early uptake, not a finding that China has matched Meta or that one agent is more capable.

A useful comparison would show which services each system can access, how often it completes the same tasks, what errors it makes and when it asks users to approve an action. Until such results are available, Muse’s reported downloads and the companies’ product descriptions show interest and design intentions, not verified rankings of agent performance.