Product1 publisher2 min readPublished
Meta's Muse stopped at the payment screen on a $7 espresso part
The Verge gave Meta's new agent a backlog of postponed chores. Muse asked for Gmail access, found the part and drafted the email, and both of the final steps, send and pay, stayed with the human.
The Product Desk · Product desk
What happened
- The Verge ran Meta's Muse through a morning of long-postponed chores, handing it new tasks while earlier ones were still running and watching it open virtual browser windows.
- Muse asked for access to the reviewer's Gmail, established that the landscaper's bid was later than the landscaper had estimated, and produced a draft the reviewer asked it to shorten before sending.
- Told to buy a replacement steam wand assembly for a Breville espresso machine, Muse located the part and added it to a cart, and the reviewer finished the order, with delivery in five to seven business days.
- The Verge rated that purchase only "Mostly" completed, and rated the third chore on the list, downgrading a security system plan, medium effort because it requires a phone call.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- contradiction The Verge's subhead credits Muse with being handy with a credit card, while its own task log shows the agent reaching the cart and the reviewer typing the card number, so the test cannot be used as proof that users will delegate payment.
- exposure The permanent action in this session was the sent email, not the order. That leaves the reviewer's negotiation with a contractor on the line for a task rated extremely low effort.
- decision Teams building agent flows have to pick, per task, which step the agent may make irreversible, because the one recorded purchase here split neatly into a recoverable search and an unrecoverable submit.
"Make it shorter," the reviewer told the agent, then read the draft again and sent it to a contractor whose price they are still waiting on [4]. The Verge wrote that "hitting send on that message made me cringe" [5], having noted a few lines earlier that "As a rule, I do not use AI writing tools" [6]. That was the lowest-effort item on the list [10], and it is the one item on that list that is now permanent.
The reversal cost is what separates these two tasks. The espresso part is about seven dollars on Breville's site, and The Verge says ordering it is as hard as typing in a card number [9]. A wrong washer costs seven dollars and a return shipment. An email in the wrong tone goes to the person who owes you a bid you have been chasing for two weeks [10], and there is no return shipment for that.
The Verge's own subhead says Muse is "handy with a credit card" [13]. The task log under it records a search and an add-to-cart, with the human completing checkout [2]. Anyone citing this test as evidence that users will hand payment authority to an agent is reading past the log. In this session the agent stopped short of buying. Who absorbs a confident wrong purchase is still an open question.
The delays are the more interesting number. The landscaper follow-up was rated extremely low effort and had sat for two weeks [10]. The seven-dollar part had sat for years [11]. Two of the three tasks in the excerpt are rated low or extremely low effort [1]. Agents get sold on hours saved. What kept these chores sitting was that they were too small to schedule and too dull to start. That is also why Muse spent the previous week pinging the reviewer about ways it could help before the reviewer sat down with it [15].
So the grid for a team shipping this is cost of the mistake against whether the mistake can be undone. The undoable boxes are where an agent can run unattended, and returns and cancellations belong there. For the boxes that cannot be undone, the confirmation step has to put the artifact in front of the user. That means the full email text and the actual cart contents, not a one-line summary of them. The Verge got that chance on both tasks and still sent an email it regretted [5].
What to watch
- Whether Meta lets Muse complete checkout itself, and what refund path it offers when the agent buys the wrong part.
- The result of the security-system downgrade task, which needs a phone call.
- Whether Muse's confirmation step shows the full draft email and cart contents or only a summary of them.