Meta's agents need to show whose interests they represent

Meta introduced Muse, its personal AI agent, on September 8. The next day, Stilla said its team was joining Meta to strengthen business AI. One announcement concerns an agent acting for a person; the other brings a team building agents for businesses into the same company. Meta's announcement, Stilla's announcement.

Those announcements do not establish that the two products already negotiate with each other, share transaction data or favour particular merchants. I would not turn common ownership into an accusation. They do make one design requirement urgent: a buyer should be able to see whose interests shaped an agent's choice.

A purchase approval is too late to answer that on its own. The user sees the selected option. The more consequential decisions may already have happened while the agent chose which sellers to consider and which trade-offs to hide.

An acceptable price can still be a poor recommendation

Meta describes Muse as able to browse and perform tasks, with user approval before actions such as purchases and sending emails. That is a useful boundary around execution. It does not, by itself, tell us how competing offers were ranked. Muse's stated controls.

Here is the hypothetical checkout I would use in a design review. I ask an agent to buy a replacement microphone below a fixed budget, delivered before a recording session. The merchant wants a sale, prefers a higher-margin product and would rather avoid a return. My priorities include compatibility, delivery confidence and a painless return if the microphone fails.

Both sides can agree on a price while disagreeing on almost everything that makes the purchase useful. An offer within budget is not proof that the buyer's agent protected the buyer's objective. Equally, a recommendation from an affiliated service is not proof that it failed. The evidence has to be in the decision.

I would ask the product to retain the alternatives considered, the constraints that ruled them out and any commercial relationship relevant to the recommendation. This is an editorial design proposal; I have not audited Muse's purchasing behaviour.

Make representation inspectable

The useful record is smaller than a complete transcript of every model thought. For that microphone purchase, it might show that two cheaper products were incompatible, one acceptable product would arrive too late, and the chosen seller offered a return policy meeting my stated requirement. A reader should be able to challenge those facts without interpreting an agent's personality.

A payment provider can constrain the amount or merchant involved in a transaction. That does not make it an independent judge of whether the shopping recommendation served the customer. Payment authorisation and recommendation quality answer different questions.

This distinction matters even when separate companies supply the agents. Common ownership makes the incentive question easier to notice; it does not create the entire problem. A supposedly independent assistant can also receive referral revenue or optimise a metric the buyer never chose.

The small experiment I would run is to change one buyer constraint while keeping the available products constant. Does the recommendation change for an intelligible reason? Can the user inspect the reason and correct a mistaken assumption before money moves?

There is no need to wait for proof of misconduct to require that interface. Useful delegation means handing over work while retaining enough evidence to judge whether the representative still represents you.

Before trusting an agent to buy for you, make it show how your priorities changed the choice.