Meta's Muse agent clicks through websites from a cloud VM that users can watch

The only account of Muse here is a dev.to post. It says it draws on Meta's Newsroom announcement, which describes a personal agent that works in the background while the user attends to other things ^[1]. Treat the details as that author's summary.

Muse drives sites the way a person does. Its model, Muse Wonder, reads screen coordinates, dynamic buttons and drop-down menus ^[4]. In our view, that means no site has to build an integration for it. In the post's example, a user asks for a car part, and the agent checks inventory at stores tied to Shopify or Walmart, compares shipping costs and prepares the cart ^[9].

Background work and a watchable browser pull against each other. The browser is visible in real time ^[3], but a task that runs while you do something else has nobody watching it. We would audit the end state: the cart, the submitted form, the confirmation.

The payment design is the best-engineered part of the account. According to the post, the gateway issues single-use virtual card numbers tied to Stripe, and neither merchants nor Meta employees see the original card data ^[5]. A single-use number cannot be replayed at a second merchant. That protects the credential. What goes in the cart is still the agent's call.

Input also arrives by email. Each account has its own address, and forwarding an electricity bill makes the agent add a Google Calendar reminder and schedule the bank payment for the due date ^[7]. In our view, any address that can trigger a payment schedule needs a sender allowlist, because forwarded mail is text written by someone else.