Meta Muse: the AI agent that shops for you but doesn't make calls

Muse, Meta's AI agent, can open browsing windows, connect to accounts, compare prices, and validate purchases once a card is linked. A usage test highlights time savings on follow-ups, shopping, and quote requests, as well as limitations, including the current inability to make calls and restrictions with certain merchants.

Online Shopping: Filled Carts, Payments via Stripe, and Confirmations

Muse can handle a substantial portion of online shopping. Once a card is linked via Stripe, approving proposed expenses becomes very simple, with details automatically filled in a virtual browser window and confirmation emails arriving afterward. The agent compared prices for espresso consumables, opting for descaling powder and water filters, and determined that Dezcal powder offered better value than the liquid version. For an old need for a spare part, Muse found the replacement steam wand, added it to the cart, and left the finalization to the user, with delivery expected within five to seven business days. In a more ambitious cart combining camping cooking gear and a propane stove, the total reached $500, demonstrating the agent's ability to aggregate purchases while remaining user-validated.

Follow-ups and Forms: Emails, Quotes, and Confirmed Appointments

For a follow-up regarding a garden renovation quote, Muse requested access to Gmail, confirmed that the delay exceeded the landscaper's initial estimate, and drafted a shortened email upon request, which led to apologies for the delay. For a tree branch removal, the agent searched for local providers, highlighted several options, collected the address, phone number, and photo, and then filled out quote request forms. A text from a company arrived quickly, and an appointment was set for the next day, for a task that had previously been postponed for a week. These cases, deemed to require low to medium effort, demonstrate the automation of preparatory steps: identifying contacts, completing forms, and initiating outreach.

Merchant Access: Amazon Blocks AI Agents, OXO and Walmart Pass

During a search for camping kitchen equipment, an attempt to access Amazon was blocked by a message indicating that AI agents violate the terms of service. However, Muse succeeded through other retailers, locating the targeted items on OXO and Walmart's websites and guiding the selection to a LavaBox stove and its case. The search, assessed as medium effort and delayed by a few days, was successfully completed despite this blockage, showing that the agent adapts to accessible channels to finalize useful steps for the user.

Phone: Subscription Downgrade Impossible to Process End-to-End

On a request to downgrade a security system, Muse was able to access the account after email verification and identified that changing the plan required a call. The agent could not make calls at the time of testing, as this capability was still under development. The procedure concluded with a call, a few exchanges with a phone robot, and then assistance from a human agent who made the change. Nevertheless, Muse streamlined the process by displaying key information such as the company's number and the names of the offers.

Usage Framework: Avatar, Browsing, and Multi-Task Execution

Muse is a consumer AI agent from Meta that operates via a customizable avatar, with suggested names like Marley and proposed appearances at initialization. In the observed usage, the agent opens virtual browsing windows, indicates its activity through typing animations, and issues prompts to showcase its capabilities. Its interest appears to be heightened when multiple tasks are assigned in parallel, which materialized through the opening of dedicated windows for each request being processed.

Data and Trust: From Gmail to Payments, a Broad Scope

Using Muse involves access to various sources and data: consulting Gmail to prepare a follow-up, collecting addresses, phone numbers, and photos to request quotes, and linking a payment method via Stripe to finalize purchases with automatic form filling and receipt of confirmations. The agent then helped to stagger expenses over the winter and flagged items likely to be on sale as the holidays approached. The test was conducted on concrete needs, from the steam wand part costing about $7 deliverable in five to seven days to a $500 cart. This scope underscores the extent of information entrusted to the agent and the necessity to rely on Meta for its processing, in a context where the user typically does not employ automated writing tools.