Your Agent's Loyalty Is a Design Choice, Not a Default

Luiza Jarovsky (https://www.luizasnewsletter.com/p/ai-agents-and-conflicts-of-interest) recommends "Designing Loyalty: AI Agents and Conflicts of Interest," which asks the question every buyer should ask before delegating: will the agent act in the user's interest when its developer's interests diverge, and who is liable if the user is harmed? The paper's answer — import the duty of loyalty from fiduciary law — is the right frame. Our position: the binding constraint is upstream of liability. A conflicted agent that picks the option slightly better for its developer produces an outcome that looks ordinary from outside, so the harm never becomes a claim. Loyalty you cannot detect is indistinguishable from loyalty you do not have. That makes logging the real battleground: as one commenter notes, EU Regulation 2024/1689 already requires high-risk AI logs to be kept at least six months. Delegation without abdication starts with a record of what the agent actually did — before any court decides who pays for it.