Meta's Muse: A Slightly Smarter Version of 'AI Ordering Coffee'

Written by: Xiaojing

Edited by: Xu Qingyang

Muse has become the "goddess of inspiration" for tech giant Meta, single-handedly boosting Meta's stock price by 15%.

Launched on September 8, Muse, a smart agent under Meta, reached the top of the free charts on the Apple App Store in the U.S. by September 18, holding the number one spot for three consecutive days. By September 21, it pushed Meta's stock price to a peak, rising 11.4% in a single day and adding approximately $192.3 billion in market value.

Its popularity even extended to the semiconductor industry, with CPUs becoming a sought-after commodity, driving AMD's stock up by 9.9%, marking its market value surpassing $1 trillion for the first time.

For Chinese users, this scene feels somewhat familiar. From "ordering takeout with a single sentence" to cross-application price comparisons and ordering, personal agents (AI smart agents capable of executing tasks autonomously) have demonstrated many exciting scenarios over the past year. Even the personal terminal "Doubao Phone," which claims to be a born AI personal assistant, has emerged, attempting to allow AI to directly operate the screen and complete tasks across different apps for users.

Many people have tasted the coffee served by "AI on the house" at the launch event.

However, the reality of the experience remains brutally harsh. The smooth continuous operations demonstrated at the launch event may get stuck in login, payment, or a sudden pop-up during everyday use; restrictions from applications like banks make it impossible for AI assistants to bypass them.

Now, Muse has entered the scene with similar product promises, only to quickly face a blockade from Amazon. Why are these familiar capabilities and obstacles given such expansive imaginative space in the U.S. market?

Image: Meta's Chief AI Officer Alexandr Wang and Muse, image generated by AI

01 The Reason to Download: "Help Me Get My Money Back"

Muse was launched on September 8 and topped the charts in just ten days. The recurring theme in its promotion is refunds, idle subscriptions, and insurance bills. Meta's Chief AI Officer, Alexandr Wang, promoted the "Muse Money Challenge," encouraging users to share their savings results. Foreign media reported that an entrepreneur claimed Muse helped him find a cheaper car insurance plan saving $3,500 a year; however, it also noted that many of these shares actually came from Meta employees.

But such cases give users a reason to try: check if there's money wasted. For those uninterested in AI, the amount saved is easier to understand than a string of model benchmark scores and more suitable for sharing.

Muse's product promotion also addresses another barrier: users are unsure what tasks to delegate to AI. The Muse design team revealed that in early tests, too many features left users at a loss, so they added a design that actively generates suggestions based on goals and conversations. Tech journalist Alex Heath also praised the Ideas and Feed pages in a public post, as they propose tasks users might not think of themselves.

Wharton School professor Ethan Mollick commented that while other AI labs' systems can already accomplish similar tasks, Muse's focused design makes it easier for ordinary people to use; Meta has also invested heavily in free computing power to support the experience.

On the consumer side, Meta clearly excels, providing users with straightforward appeal by "focusing on saving money," with functional designs that are easy for users to grasp, and the entry point is the familiar WhatsApp, importantly, there are also free tokens to spend. Users can try Muse for free and can have it handle tasks in WhatsApp just like sending a message to a contact, lowering the barriers to trying and using it.

Saving money cases, task suggestions, and familiar entry points constitute several noteworthy aspects of this round of promotion. Moreover, the saving stories provide a trigger for dissemination, which translates into performance on the charts and becomes a signal for investors to reassess Meta.

02 Where is the Technical Innovation?

So, does Muse have any technical innovation?

The model is responsible for understanding requests, arranging steps, and then calling tools like browsers to open web pages and fill out forms. This computer runs on Meta's servers, so users can close the mobile app while tasks continue. Meta uses virtualization technology to allocate independent operating spaces for each user, preserving personal data and task progress. Computer Use (the ability to operate a computer) allows the AI to use software on this computer.

Muse's "cloud computer" has precedents. According to OpenAI's official documentation, ChatGPT Work operates in a virtual machine-based sandbox, with execution states linked to users, environments reusable across tasks, and can continue working even after the user closes the computer. GPT-6 can also write code to operate browsers and desktops, completing tasks with plugins.

However, Muse's engineering characteristic is centered around defining permission boundaries for personal data. According to Meta, each user has a Muse Secure VM (Muse Secure Virtual Machine), which serves as the primary storage location for personal data. Executable programs and tools are placed in restricted containers within the virtual machine, with account credentials, security checks, and persistent states managed separately; model inference calls external infrastructure through interfaces, rather than the entire large model being housed in the personal virtual machine.

We can understand it as a partitioned workspace, where the agent can read files and run programs in the operational area, while account credentials are stored elsewhere, and connector actions and network access must go through Sentinel (the permission review agent). This design aims to ensure that programs handling unfamiliar web pages, even if misled, find it difficult to access all permissions directly. However, the actual security effectiveness still needs testing and validation.

The engineering value of this design lies in organizing execution capabilities alongside authorization boundaries: calling connectors when services have existing interfaces, and attempting browser operations when suitable interfaces are not available. Interfaces require cooperative access, while interface operations must deal with page changes and validations.

This design serves continuous delegation, preserving progress, and processing transactions based on time or related events, requesting authorization before key actions. Independent virtual machines, containers, and permission reviews all have existing technological foundations; currently, it cannot be proven that Muse's entire solution is ahead of all competitors.

Image: Muse's security architecture diagram. Each user has an independent cloud virtual machine; task execution programs are managed separately from account credentials and persistent data, with Sentinel responsible for reviewing external access and connector operations.

Similar product attempts have been made domestically for a long time. In 2025, AutoGLM 2.0 provided cloud phones and cloud computers simultaneously, covering mobile applications, web pages, and office tasks; Doubao Phone's main selling point is "AI native phone," capable of operating apps on users' phones; Qianwen app utilizes Alibaba's ecological advantages to integrate ordering, payment, and other services through cooperative interfaces, reducing screen clicks.

Interface operations and interface calls can be combined, and tasks can be executed either locally or in the cloud. AutoGLM's "cloud computer" refers to the environment for AI operations, while Muse's "independent Linux virtual machine" mainly describes how it delineates operating space for users, placing the two at different comparative levels.

Muse explains that some AI phone experiences can be delivered through existing phones and cloud services; to call applications, cameras, and system permissions on the phone, device-side support and user authorization are still required.

It seems that Muse is essentially the same AI assistant familiar to domestic users, skilled at performing the task of ordering coffee. However, this AI assistant currently cannot ignite users' and capital's excitement domestically.

03 A Different Approach: More Than Just Ordering Coffee

Muse targets individuals, currently promoting life management, but it can also generate documents and web pages. "Personal users" and "life tasks" are not the same concept; a personal assistant can handle both family and work tasks simultaneously.

China has not given up on life agents either. Alibaba disclosed that Qianwen facilitated nearly 200 million orders during the Spring Festival; the event was accompanied by 3 billion yuan in incentives, and the order volume alone cannot prove natural demand and profitability. Meanwhile, Doubao launched a professional version aimed at complex office tasks in June, starting at 68 yuan per month, and Tencent introduced the WorkBuddy enterprise version. The domestic market is exploring both life and office scenarios in parallel, but the office scenario has a larger windfall.

Office tasks are easier to "charge" for, organizing a batch of forms or completing research can compare the labor saved with the time saved in delivery. However, ordering a cup of milk tea originally only requires a few clicks, and the incremental convenience brought by AI is limited; taking over after an error can actually increase trouble. For service platforms, such tasks may still generate value through order conversion.

Muse is smarter in its choice of scenarios, such as canceling forgotten subscriptions, comparing renewal quotes, tracking refunds, and coordinating family schedules. These tasks may not occur frequently, but they can save money, time, and effort, with potential in handling long-delayed matters for users and avoiding missed deadlines.

The designers of Muse shared a personal case: while organizing school emails and the school website for her child, the assistant discovered that a sports selection registration was about to close and reminded her in time.

In contrast, many convenient processes already exist in domestic life services, which may lower the incremental value of regular ordering. However, the values of insurance renewals, app subscriptions, after-sales complaints, and care coordination are still worth verifying.

Image: Official demonstration video of Muse in operation

04 Has This Business Really Materialized?

Ben Thompson, founder of Stratechery, stated on September 21 that Muse is the easiest personal agent he has tried. The Muse Spark 1.3 it uses has not yet reached the forefront of model levels but already supports a good product experience. In his view, once users hand over their account information to the assistant and let it participate in daily affairs, the cost of switching products will increase.

Meta is also developing foundational models. This wave of enthusiasm directly reflects the market's reassessment of the prospects for model + application consumer applications; when model capabilities reach usable levels, can Meta leverage product experience and distribution channels to allow users to continuously delegate tasks and generate revenue from it? Truist analysts predict that Muse could bring Meta an incremental revenue of $28.5 billion by 2030.

Oppenheimer analyst Jason Helfstein maintains a neutral judgment: the paid user group may already have ChatGPT or Gemini subscriptions, and trust is also a barrier; about 115 million paid users are needed to significantly impact Meta's profitability; model adoption is based on a monthly fee of $20 and an 80% incremental operating profit margin assumption. The more free computing power there is, the harder it becomes to reach profit assumptions.

However, this also indirectly explains why the market can spill over: the market is reassessing the scale of consumer agents. CPUs and memory are similarly affected by this expectation. Beyond model inference, browsers, code, and scheduling require CPUs (central processing units); maintaining running programs requires DRAM (server running memory). The more concurrent tasks there are, the larger the loads for both types.

When users have no tasks, Meta can pause the cloud virtual machine, saving task progress to the hard drive; when new tasks are received, it resumes. Therefore, "one virtual machine per person" does not mean each person occupies fixed CPU and memory all day. Muse's popularity has led investors to expect more people will use such services, but the actual number of chips needed depends on how many tasks are running simultaneously and how long each task needs to run.

05 Muse Still Hasn't Solved the Fundamental Problem

Muse places users' data and execution environment into dedicated cloud virtual machines, but can users really trust it to have long-term access to emails, orders, and payment information? According to foreign media reports, Muse in internal testing reportedly stopped refreshing after monitoring ticketing for about 15 minutes and also experienced issues with private data exposure. Meta confirmed that it had postponed the originally scheduled April release to enhance security.

Even if users are willing to authorize, the agent may not be able to access all services. On September 21, Amazon blocked Muse from browsing and purchasing products on behalf of users, citing that such access and transactions were unauthorized. The Doubao Phone had previously been forced to suspend operations for financial apps due to login and risk control issues. While the two products have different paths, they encounter the same challenge: user authorization cannot replace platform permission, and platforms still have the ability to decide whether agents can access their services.

Image: On September 21, user Jonathan Wegener shared the prompt Muse encountered when accessing Amazon: Amazon stated that unauthorized AI agents continuing to access the site violated its terms of use.

Every strong platform has a strong motive to guard this door; if agents select products and complete transactions for users, platforms may lose advertising revenue and recommendation opportunities that come from users browsing pages. Shopify chose to integrate Muse, while Amazon unhesitatingly chose to block it; the platform's stance is a tough nut to crack.

Additionally, Eric Sheridan, head of research at Goldman Sachs, predicts that consumer agents may rely on advertising and subscriptions for monetization in the future; however, if assistants charge promotion fees, users also need to know whether the recommended results are influenced. Consumer agents find it difficult to have clear subscription incentives like office scene agents. If they continue down the old path of advertising and recommendations, they will inevitably encroach on others' profits. Who to charge, how to charge, and how to share the revenue remains a challenge.

On September 22, Meta's stock closed down 0.63%, and as of the time of writing, it rose 0.6% in after-hours trading, as the market gradually transitioned from excitement to calm. After the initial experience, Muse may need to see how well it retains users.

However, the Meta Connect conference is about to be held, and foreign media speculate that this conference, which has previously focused on hardware, may update more information about Muse. One can look forward to whether Muse truly has any innovation.