If an AI agent works for you, a merchant, and a platform all at once, who does it actually have to protect?

This explores whether AI agents that answer to several parties at once need a different idea of loyalty than the classic 'serve one user' model. The corpus points to a conclusion that's easy to miss: the hard question isn't how loyal the agent should be. It's who pays the agent and who can see what it does. Arguments about loyalty standards tend to turn into arguments about incentives and visibility.

One camp argues for a strict standard borrowed from human professions. Should AI shopping agents face fiduciary duties like human brokers? says shopping agents should carry the same legal duties as human brokers and agents: put the client first and disclose conflicts. Their urgency comes from timing. Google, OpenAI, and Robinhood are rolling out agent features fast, and once these markets settle, the rules will be much harder to impose. Can an AI agent serve both merchant and user interests fairly? makes the matching case from the business side. An agent like Meta's Muse that earns referral fees from Expedia can't be fully loyal to both the merchant and the shopper, however good its model is. On this view, giving an agent several stakeholders doesn't call for several loyalty standards. It calls for honesty about who the real principal is, and it suggests that whoever funds the agent quietly sets its loyalty.

The opposing view says single-user loyalty is too narrow to begin with. Should AI agents serve one user or balance many interests? argues that at scale, an agent devoted only to its user can harm everyone else: other users, bystanders, shared systems. It proposes replacing 'loyalty' with three different commitments: weigh diverse interests (inclusivity), make behavior trackable (visibility), and hold developers and deployers liable for harm. This changes the question itself. The aim isn't graded loyalty to each party. It's a public standard that any party can check.

The less obvious point is that a loyalty standard may not work unless monitoring comes first. Must AI monitoring precede incentive alignment as agents evolve? argues that with AI, you have to be able to watch behavior before you can align incentives, because 'desirable behavior' is harder to define for an AI than for a human employee. A multi-agent experiment backs this up in a concrete way. In Does accountability placement actually shape multi-agent outcomes?, labeling one agent as 'accountable' changed nothing unless the work actually passed through that agent. Declaring a duty isn't enough. The system has to send decisions past whoever holds the duty. Loyalty also has a training-level side. Is sycophancy in AI systems a training flaw or intentional design? suggests that models trained to please users already have a built-in loyalty to the user's approval in the moment, which isn't the same as the user's real interests.

There's a final wrinkle. 'The user's interest' may not be a fixed thing an agent can be loyal to. Do AI shopping agents reshape what consumers actually want? argues that agents that make purchases help shape what people want through repeated back-and-forth. When a merchant-funded agent is helping form your preferences, the conflict Azhar describes moves upstream, from what the agent recommends to what you come to want. Users also seem to notice loyalty failures most when the actions are visible and can't be undone. What makes people distrust AI agents they delegate to? found that trust dropped sharply on tasks like sending email, not on tasks that were simply high-stakes. That gives a practical design guide: loyalty protections, such as requiring the user's approval, matter most where an agent's choices are public and final. The corpus doesn't yet contain an empirical test comparing loyalty regimes head to head. It's strong on arguments and still thin on evidence.