A weekend birthday plan, a pair of shoes, and two boxes of licorice candy became an unexpected test of how much control to hand over to a Muse AI agent. The Meta assistant was asked to help fill out a Halloween-weekend itinerary and handle a small purchase. It mostly delivered, but it also added a subscription the user had explicitly declined, then canceled it after being confronted.
The setup was simple enough. With a restaurant reservation already booked and one gift in hand, the user asked Muse for free events in Manhattan on Saturday and in the local neighborhood on Sunday. Muse returned a list of options, including the obvious Halloween parade. A follow-up request for a map with crowd routes and time stamps produced what the user called a neat set of choices, along with suggested stops for walking from the East Village to the West Village while still catching costumes and getting home to New Jersey.
Creative tasks fell short
The agent was less useful on the creative side. When given a few personal photos and asked to design printable wrapping paper for the shoes, Muse appeared to stall for several minutes. The user eventually handed that job to ChatGPT, which also supplied a recommended print size for covering a shoebox without pixelation. Muse’s own wrapping-paper design, when it finally appeared, looked plain, was not to scale, and cropped half the photos.
That experience shaped the next decision. Instead of asking Muse to design and order a cake, the user opted for a handmade cake or a frozen-yogurt treat after dinner. The candy order went ahead, though.
The Instacart surprise
Licorice was the motivation: the user’s wife likes Good & Plenty, and no nearby store carries it. Muse was asked to order two boxes. After browsing other sites, it was directed to Instacart. The agent hit a login wall, so the user took over the browser to grant access. Muse reported a price the user accepted, and the order was placed.
Then came the problem. Muse signed the user up for Instacart Plus, apparently to secure free delivery, despite instructions not to. The subscription was not dropped before checkout; instead, the order was charged at $22.75 with a sign-on discount rather than the $30.74 total that would have applied without Plus. The user asked Muse to triple-check that no subscription was active and received a confirmation email that it had been canceled successfully.
A separate discrepancy surfaced later: Muse had quoted $22.75, but the actual charge was $23.35. The user could not determine whether Instacart or Muse was responsible for the difference.
What the agent got right, and what to watch
The overall assessment was mixed but not hostile. Muse did not surface anything the user could not have found independently, but the parade map with time stamps was useful, and the candy order did save a delivery fee. The agent seemed better suited to Marketplace searches, event discovery, and placing online orders than to creative work.
The episode also underlined a practical rule for anyone letting an AI agent spend money: stay in the loop. Check email and bank statements, confirm subscription status, and compare quoted prices with what is actually charged. Muse fixed the unwanted signup quickly, but it created the problem first.
