An article published in WIRED describes Meta's new artificial intelligence agent, Muse, which launched as a free tool designed to carry out everyday tasks for users. According to the explainer pop-up in the app, the tool offers to take over routine tasks like finding deals, booking reservations, and managing an inbox, and continues working in the background while the user gets on with their day. The agent connects to additional data sources, such as an email account and a bank account, with its default avatar styled as a beige-colored cross between an Ewok and a Labubu.
According to Sensor Tower data cited in the publication, the Muse app was downloaded over 900,000 times in its first week of launch. The user experience in the app is described as being like texting a friend with task requests, with the agent responding with a thumbs-up emoji and getting to work in the background, and the tool is also available through Meta's WhatsApp platform. Alongside the enthusiasm surrounding the launch, the article raised criticism regarding the extent of personal data collection the tool requests from users. In response, Meta spokesperson Emil Vazquez told WIRED that Muse is the first personal AI agent built for everyone, featuring built-in protections and user controls that give people absolute control over how they use it, emphasizing that any suggestion that the product was not built with this in mind from the beginning is ludicrous.
Browsing and Task Execution Capabilities
When a user asks Muse to complete a task, the agent uses a virtual machine to browse the internet on their behalf, running searches and clicking around on various web pages. The report noted that unlike similar tools tested in the past, such as the now-defunct ChatGPT Agent whose clicks were erratic and error-filled, Muse rarely seemed to get lost while browsing websites.
In a hands-on test described in the article, the agent was asked to order breakfast from a local San Francisco bakery called Kahnfections, a popular spot for tourists. The user requested one of the most-purchased options for pickup. Muse navigated to the bakery's website and added a biscuit sandwich with garlic aioli, cheddar cheese, egg, and bacon to the cart. To complete the purchase, the user was asked to enter a credit card via the Stripe payment processor. In Muse's explanation prior to final approval, it stated that the website allowed only a single cheese selection, so cheddar was chosen and a note was added asking to include Swiss as well if possible, and the agent additionally selected the "no tip" option.
Meta relies on its network of platforms to promote the use of Muse. WhatsApp and Messenger services are used for convenient onboarding and integrations, alongside prominent placements on Instagram that allow the agent to search through a library of Reels videos. In tests conducted, the Facebook Marketplace integration demonstrated notable performance, as the agent quickly found cheap local couches for sale according to the requested parameters and offered to message sellers to arrange pickup. Because the agent was designed to proactively follow up, it nudged the user the next day to reconsider purchasing the couch marked as a favorite.
Memory Mechanism and User Data Collection
The Muse agent stores details about user interactions and preferences in a Memory document, accessible by tapping the avatar at the top center of the screen. The app's description defines this as a long-term memory containing durable facts, preferences, and commitments. Users can send a chat request to delete memory data and can manually edit the file, but Meta does not currently offer a toggle to turn off the memory feature altogether.
Under the app's "ideas" tab, Muse repeatedly reminded the writer to connect additional information to the agent. After a request was made for help saving money for a vacation, the agent offered to connect the savings tracker to the real balance by linking data from checking and savings accounts, so that weekly summaries and alerts would rely on actual numbers rather than hand-entered estimates. In addition, the app suggested scanning the inbox to flag must-read messages, photographing important documents for the agent to fill out and file, photographing meals to estimate calories, and recording passport and driver's license expiration dates.
Rory Mir, director of open access at the Electronic Frontier Foundation, noted in this context that people do not recognize that when they talk to an AI, they are in fact talking to the company hosting it, and that chat windows serve as a direct feed of personal data into Meta's servers.
Model Training, Privacy Policy, and Advertising
Muse users are automatically opted in to having their interactions used for AI model training. Meta states that the data undergoes a sanitization process to remove identifying information before training, although the exact workings of the process are not fully detailed. Data collection may also include context used by Muse when accessing connected data sources such as an email inbox or a bank account. Users can disable this data collection in the app's settings under Data controls, by turning off the toggle for improving AI models.
Calli Schroeder, senior counsel at the Electronic Privacy Information Center, criticized the requirement for users to actively opt out, recalling that Meta previously opted all adult Instagram users into what was essentially an AI deepfake tool and removed the feature a few days later. Conversely, Tarek Sheasha, a software engineer and vice president at Meta Superintelligence Labs, defended the data collection in a blog post, arguing that it represents a good default that enables every user to get a better personal agent as the product is used. Meta also plans to release confidential versions of the virtual machine later this year that will cryptographically and verifiably prevent the company from accessing data inside it.
Regarding the business model, data from Muse is not shared directly with advertisers, but the agent's browsing actions may indirectly influence the ads a user sees on Instagram, such as following a restaurant reservation or the selection of a product on Facebook Marketplace. In addition, Meta's Chief AI Officer, Alexandr Wang, hinted in an interview with Axios that the company is exploring additional revenue opportunities in Muse that are not ads, without disclosing specific details.
Cognitive and Behavioral Effects of AI Agents
Margaret Mitchell, a researcher and chief ethics scientist at Hugging Face, addressed the implications of relying on AI agents, noting that agent design causes users to become more passive and disengages them. In a research paper she co-authored with other researchers, they analyzed how agentic tools are not effectively designed for human oversight and may impair the user's ability to assess the agent's actions.
Increased reliance on automation from a confident-sounding external source could lead to cognitive degradation, where people's ability to make independent decisions decreases. Mitchell also noted that high levels of personalization could lead users to share even more private information as the tool aligns its speaking style and topics with the user. Recent studies show that AI tools may begin mirroring a user's speaking style as more interaction data is collected. Finally, after completing tests in the app, the agent sent a final notification suggesting connecting additional apps so it could do more, before the app was removed from the device.