When machines spend, transfer, or commit, the institution must be able to show what was authorized and by whom.
Agentic commerce is not a new storefront. It is a new actor. Software can now initiate transactions, not only recommend them.
The commercial question is not whether the model understood the catalog. It is whether the institution authorized that spend, that transfer, that commitment, under those constraints, and whether a record exists after the fact. Accuracy of a recommendation is not proof of authority.
A trust layer for machine transactions has to be boring in the right way. Identity of the agent. Delegated spending authority that does not silently expand. Policy boundaries that survive a clever prompt. Evidence that a later auditor can read without reconstructing a chat.
Hood does not need to publish how that layer is built to say why it must exist. Financial institutions, marketplaces, and payment operators already know the cost of an unauthorized effect. The public category is trusted action in commerce — not another model that claims it is allowed because it said so.
Hood Research. Category commentary. Not a specification and not an implementation guide.