Meta Muse downloads don’t tell users whether its agent will get tasks right

The South China Morning Post reported that Muse passed 2.5 million downloads in its first two weeks and topped the free-app rankings in Apple’s and Google’s U.S. stores. That figure signals early interest, but it does not establish how many people kept using Muse, whether it completed tasks correctly or how much time it saved. (The South China Morning Post’s report)

The Tom’s Guide reviewer’s report of 10 tasks offers a closer look at possible uses, including reminders and alerts. But one person’s test is not a controlled comparison or a measure of how the product performs for other users. A useful evaluation would check whether reminders arrive when needed, orders match instructions, and users can inspect and correct mistakes. (The reviewer’s account)

The agents do not share a playing field

Tencent’s Xiaowei takes a different path by operating inside WeChat, where it can reach services already in the app. Quartz reported on July 3 that Tencent was testing it with a small number of users. The assistant could respond to voice or text and open mini-programs; Tencent had not set out the full range of actions in the trial. (Quartz’s report)

On Wednesday, Semafor reported that Xiaowei could place orders through businesses on WeChat, but that early tests also produced mistakes. That reporting establishes a reported capability and early errors, not consistent performance or general availability. Quartz also reported agent trials by Ant Group inside Alipay, including evaluations of ride booking and food ordering; it did not report a broad public launch.

Integration may make some tasks easier to start because users can act through services already inside an app. But it also limits the agent to services available there. Quartz reported that WeChat and Weixin had more than 1.4 billion combined monthly active users, citing Tencent; that figure describes the apps’ audience, not Xiaowei’s reach. Distribution can make an agent easier to access, but it does not establish that the agent completes tasks reliably.

An agent that works across outside websites faces a different obstacle: Semafor reported that Amazon had blocked Muse. Neither route is a clean measure of who is ahead; each exposes users to different constraints. (Semafor’s account of the US and Chinese approaches)

OpenAI’s dots add another stage to the picture. The company said on Sept. 29 that rollout was beginning for selected paid plans in eligible markets, with tasks running in the background through connected apps. That is a staged rollout, not general access for everyone, and the announcement alone does not establish how well dots handle real tasks. (OpenAI’s announcement)

Users need control, including non-users

A personal agent may act on information about more than the person who installed it. Tom’s Guide reported that independent AI safety and security researcher Karan Joshi extracted Muse instructions that could organize information about friends, family and colleagues. The report does not establish that every such person has a profile, or that any profile is complete or accurate. It does raise a practical issue: people who never signed up may have no clear route to correct a mistaken detail recorded about them. (Tom’s Guide’s report on the extracted instructions)

That issue belongs alongside task success, not behind a race for downloads. An agent that saves a user steps but makes consequential actions hard to review, or draws in other people’s information without a way to correct it, shifts costs onto people who may not have chosen to use it. The available reports do not show how often these systems make mistakes, how much time they save across users or whether they have changed work and pay.

Companies control the first safeguards: permissions should be narrow and revocable, consequential actions should require clear confirmation, and users should have an inspectable record of what the agent did. People described in an agent’s memory should also have a route to seek correction or deletion. Regulators can decide whether those rights must be required. The test of personal AI should be whether it helps people complete tasks on terms they can understand and control—not whether its maker appears to be winning.