Start with what the merchant's server sees when this thing arrives. On the described architecture the agent carries its own computing environment rather than borrowing a person's browser [1], so a checkout at one of the five sites the prototypes were trained against [3] is a session no hand on a mouse initiated, and the only thing separating it from unwanted automation is whatever identity Meta chooses to attach to the request. Nothing in the account describes such a signal; it lists mistaken purchases, privacy and third-party site policies as open questions and leaves them there [9]. Count the outside companies named across the prototype targets and the service connectors and you get seven [2]. That none of them is reported to have agreed to anything [9] is itself the finding: authentication of a delegated buyer hasn't been settled with a single party the agent would actually visit.
The pricing term is the part I would read twice. The reported high-end tier of about $200 a month comes attached to higher usage limits [6], which tells you the meter runs on work performed rather than on a seat occupied, and $200 a month is $2,400 a year [1] for errands a household currently runs itself. That is a subscription business, or rather it is the version of Meta's AI spend that carries a price facing the customer instead of the advertiser, and note where it points the product: planning inside Meta AI and a Business Agent handling merchant workflows on WhatsApp and Instagram [7] sit on both ends of their own transactions, while this agent walks out to other people's checkouts, where the only reported monetisation is the subscription [6].
This is probably wrong, but the binding constraint over the next year looks less like model quality and more like merchant authentication: whether a store can tell a delegated buyer from automation it would rather refuse, and whether it would sooner price that distinction than block it. The counter-case is decent. A bot that completes a checkout is a paying customer, and a merchant turning it away is turning away revenue, so the settlement may arrive as signed agent identity and partnership terms, in which case the question gets answered commercially before it becomes a public fight. The third path is that none of this ships in this shape, since timing and final pricing are unconfirmed and a Meta spokesperson declined to comment [8], and a tool that only reached employees outside the core lab last month [2] can stay internal for a long while.
If Meta launches with named merchant agreements and an agent identity scheme already in place, the pricing of delegated traffic will have been settled privately by the parties with the most to lose, before anyone outside gets a vote. Absent that, the reported approval gate on payments and bookings [5] is worth reading for the scope of what it covers: it keeps a human liable for the charge, but the store on the other end of that charge still gets no signal telling it who, or what, just bought something.
