Meta's AI agent Muse helps with chores but makes spending money dangerously easy
A Verge writer tested Meta's new AI agent Muse, which handled everyday errands from emails to online shopping. The agent completed several long-delayed tasks but also made it strikingly easy to spend money.
Meta has released an AI agent called Muse, designed to handle everyday tasks on a user's behalf. A Verge writer tested it under the name Marley, assigning it several chores she had put off for weeks, months or even years.
Tasks handled
Marley was asked to follow up with a landscaping company about a delayed renovation bid. After being granted access to the writer's Gmail, it drafted an email and shortened it on request before it was sent. Another task involved ordering a replacement steam wand part for an espresso machine — Muse found the part and added it to a shopping cart, though the order still had to be finalized manually.
When asked to downgrade a home security subscription, the agent discovered that a phone call was required — a capability Muse doesn't yet have, though Meta reportedly is working on it. For a more successful task, the agent searched for junk removal companies to deal with a fallen tree branch, using the writer's address, phone number and a photo to request quotes, resulting in a scheduled pickup.
Shopping and privacy concerns
Muse was most active helping with shopping — assembling a camping kitchen setup and a propane fire pit priced around $500 total, along with espresso machine supplies. The agent attempted to use Amazon but was blocked for violating the platform's terms of service, so it purchased items through OXO and Walmart instead. After linking a credit card via Stripe, approving purchases became remarkably easy.
The writer noted that within 40 minutes of use, she had shared her full name, address, credit card number, a photo of her backyard, and login credentials for her home security account. Meta states this information is stored securely and that Marley itself cannot access it. The writer concluded that while the agent resolved several outstanding tasks, it didn't ultimately free up meaningful time — new tasks kept emerging, and time was instead spent conversing with the agent.
