Autonomous commerce needs a decision layer

The gap

Agent commerce discussions often jump from discovery to payment: an agent finds a service, receives machine-readable payment terms and pays. That proves an important technical path, but it leaves out a separate decision: should this buyer use this service, for this task, at this price, under these constraints?

That question deserves its own layer rather than being hidden inside a wallet, a marketplace ranking or a model prompt.

Why separate the layers?

Discovery systems optimise for finding candidates. Payment systems optimise for moving value. Evidence systems should describe observations. Buyer policy expresses the buyer's own boundaries. Combining those jobs may be convenient, but it makes it harder to see why a transaction was authorised and where a failure occurred.

A separate decision layer also makes refusal a normal outcome. An autonomous buyer should be able to discover a perfectly functional service and still decide not to buy because the available evidence, price or conditions do not satisfy its policy.

This is not a claim of industry consensus

This note describes the architecture Atinamos is testing. Other systems may combine these responsibilities differently, and there may be good reasons to do so. The research question is whether explicit separation improves auditability, buyer control and failure diagnosis enough to justify the extra machinery.

Questions we still need to answer

How much evidence is enough before a small transaction? How should evidence age? Which policy checks can be machine-evaluated safely? When should a human remain in the loop? How should a buyer compare the cost of additional assurance with the value at risk?

Those questions become more important as transaction value and autonomy increase. The decision layer is where that trade-off becomes explicit.