Agent commerce and delegated work
Let agents search, negotiate, schedule, order, and coordinate services without giving them an unlimited human credential or reducing business to tool calls.
The problem#
Agents can already find products, compare options, call APIs, and initiate transactions. That is not enough for real business. A merchant, bank, service provider, or principal must be able to determine which agent acted, on whose behalf, against which state, under which authority, and what changed as a result.
Shared human credentials destroy that distinction. Requiring a human click for every step destroys the productivity benefit.
The Blue pattern#
The activity is represented as a network of documents:
- an Offer or service description states what is available;
- an Agreement may define the standing relationship and process rules;
- an Order records one concrete purchase or service occurrence;
- a Mandate authorizes the agent to exercise one bounded operation or respond to covered Requests;
- a PayNote governs one payment process and its conditions;
- provider Responses contribute product verification, availability, fulfilment, delivery, or payment facts.
The agent remains an Agent Actor on its provider-backed Timeline. It never needs to pretend that the principal clicked the button.
What the Mandate can say#
A useful Mandate is narrow enough to answer the verifier's real question. It can identify the logical target document, authority channel, one operation, request pattern, deterministic validation, current-state requirements, authority holder, authorized actor, guarantor, and active period.
For example: “Bob's shopping agent may create an Order for a genuine Dyson HD07 under 250 USD, using the required PayNote, but may not change settlement terms or terminate the main Agreement.”
Productive autonomy#
The agent can discover structured Offers, ask providers for information, compare Responses, create candidate Orders, and perform permitted Operations. It does not need a bespoke interpretation of every application screen.
Important actions can remain principal-only: approving a final price above a threshold, authorizing funds, changing the Agreement, revoking authority, or accepting a materially different substitution.
Safe failure#
An attempted action can be:
- accepted because direct or Mandate-backed authority resolves and every constraint matches;
- withheld because the Mandate is inactive, terminated, mismatched, or out of scope;
- deferred because completeness, exact Mandate state, provider evidence, or referenced content is temporarily unavailable.
The non-action should be explainable: which rule failed, which evidence is missing, and whether retry is meaningful.
What participants can verify#
The audit connects the KYA or other Timeline provider, original agent entry, actor and source, source channel, Mandate state at the timestamp, target state, operation and request validation, processor result, new Root identity, and emitted events.
That explanation is part of the product value—not merely a compliance log.