Meta's Muse agent booked a Santa Monica hotel room twice after misreporting the payment

The failure sits between the payment and the record of the payment. Muse put the card through a process the Benzinga account describes as circuitous, running through both Chase and Stripe. Then it returned an error saying it had not completed the transaction, that it did not know why, and that the card had not been charged ^[3][4]. The charge had gone through ^[5].

The correction is the more instructive part. The reviewer sent Muse a screenshot of the charge and told it the payment had gone through, and Muse came back with a Marriott confirmation number without being asked to ^[6]. Two rooms were waiting at the front desk ^[7]. The screenshot did not reconcile the booking that already existed. It produced a second one.

The $408 covered the extra two nights before taxes and fees, about $204 a night ^[8][9]. That is small on a corporate card. The recurring expense is the detection and dispute work. Someone has to distrust the agent's confirmation, check the statement, and obtain the reversal, and here the reversal came from a front-desk clerk who canceled the duplicate ^[8].

This is one reviewer's few days with a consumer app, not a controlled deployment. The account covers a handful of tasks by one person and does not report how often the payment handoff breaks ^[18]. The defect is not about how good the model is. If the booking call has no idempotency key and no read-back of payment state, a retry after an ambiguous error creates a second reservation whichever model is driving.

"I've found interacting with the AI something like trying to wrangle a lackluster employee," the reviewer wrote ^[13]. The account also describes setting up Muse on a work laptop with a phone number, then being unable to get in on a Mac mini or an iPhone because the login codes sent by text returned error messages. He eventually created a second account linked to Facebook, which was itself tied to the same phone number ^[10]. Muse did complete a Resy reservation and schedule an Uber ^[11]. "I can't truthfully tell you it was faster or easier than if I'd just gone directly to those apps," the reviewer wrote ^[12].

Capital for this category is moving faster than the reliability of it. Instinct, a personal-agent startup founded last year, has been seeking funding at a valuation of around $10 billion, four times the $2.5 billion it was valued at a month earlier ^[14][15]. It is not widely available while it works through a compute shortage ^[16]. Meta is aiming Muse at the same everyday people who use Instagram and Facebook in the billions ^[17].