Technical Foundation

Reference Architecture_

The technical foundation for agentic commerce — logical architecture, deployment models, and system interactions.

Logical Architecture

This reference architecture is designed for autonomous economic execution across the lifecycle of physical goods.

The blueprint is organised into four horizontal layers and three vertical concerns. The purpose is not abstraction for its own sake — it is to create durable contracts between systems so agents can execute safely across enterprises, partners, and physical networks.

Figure 1: Logical architecture showing four horizontal layers and three cross-cutting concerns

Horizontal Layers

LayerPurposeWhat it owns (contracts)
ExperienceHuman + agent entry pointsConsent, mandate UX, agent APIs, real-time state and exception handling
OrchestrationExecution under policyWorkflows, tool use, approvals, retries, dispute triggers, partner coordination
DomainEconomic primitives for physical goodsProduct identity/state, value & routing, lifecycle actions, payment intents and settlement semantics
FoundationRuntime for scale and trustEventing, storage, security, observability, tenancy, key management

Vertical Concerns (apply to every layer)

  • Trust & Identity — verifiable principals (users/orgs/agents), delegation chains, revocation, credential hygiene
  • Governance & Policy — deterministic constraints, approvals, dispute handling, compliance evidence
  • Observability & Provenance — auditable linkage between decisions, mandates, state transitions, and settlement events

Non-Negotiable Invariants

Agents may propose actions, but they cannot execute without (1) identity, (2) a mandate, (3) policy evaluation, and (4) provenance. These invariants turn autonomy into trustworthy execution.


Payments Domain

Agentic Payments is Execution Semantics

Agentic commerce requires payments designed for delegated authority, multi-step execution, and lifecycle-linked settlement. This is not a checkout optimisation. It is the economic layer of autonomy.

The Payments domain provides the primitives that allow agents to transact on behalf of a principal across both sides of the checkout "beep" — including refunds, credits, split payouts, and outcome-based settlement.

Domain Responsibilities (canonical primitives)

  • Payment Intents — structured execution instructions (purchase, subscription, refund, credit, payout, split settlement)
  • Mandates & Policy — delegated authority with constraints (limits, categories, approvals, time windows, merchant allowlists)
  • Tokenisation & Credentials — policy-bound credentials suitable for machine-initiated execution
  • Progressive Authorisation — hold, adjust, capture, reverse across multi-step flows
  • Settlement Semantics — partial refunds, credits, charge adjustments, split payouts, escrow/holdbacks tied to verified outcomes
  • Ledger & Provenance — immutable linkage between product state transitions and money movement for audit and dispute handling

What Changes in an Agentic World

  • Authorisation becomes a process, not a moment
  • Settlement follows outcomes, not clicks
  • Refunds and credits become programmable, triggered by inspection/verification events
  • Payouts become multi-party, supporting marketplaces, service networks, and recovery partners
  • Mandates become the control plane for economic autonomy

Integration Points

Card Networks & Token Services

Tokenised credentials, delegated authority patterns, consumer-grade acceptance and dispute flows.

Account-to-Account (Open Banking)

Direct settlement and programmable mandates for certain geographies and B2B execution scenarios.

Merchant / PSP Platforms

Integration into acquiring stacks for partial capture, refunds, reconciliation, and settlement reporting.

Financing & Protection

Optional financing, warranty, and insurance primitives triggered by product state and risk signals.


Physical Deployment

Autonomous economic execution is an edge ↔ regional ↔ data centre system. Physical commerce is real time, distributed, and operationally unforgiving.

The blueprint is designed for hybrid deployment. Placement is driven by latency, resiliency, data sovereignty, and the physics of interaction.

Figure 2: Tiered deployment model (edge ↔ regional ↔ cloud) aligned to latency and sovereignty

Deployment Tiers

TierTypical PlacementPrimary WorkloadsResiliency Goal
EdgeStore, kiosk, warehouseIdentification, authentication, local gating, offline-safe executionContinue critical flows during connectivity loss
RegionalGeo cluster / regionRouting, risk scoring, inventory sync, partner orchestrationPredictable latency and controlled blast radius
Cloud / AI FactoryGlobal control planeContinuous optimisation, training, simulation, governance, policy distributionMulti-tenant scale and coordination

Why NVIDIA is Structural Here

This industry requires:

  • edge inference (vision, identity, classification) at points of interaction
  • predictable regional execution for routing and risk controls
  • AI factory workloads for continuous optimisation, simulation, and governance at scale

Compute is not "supporting" commerce here — it becomes the substrate of economic execution across the physical world.

Edge Requirements

For in-venue deployments, edge nodes should support:

  • Intermittent Connectivity — core flows degrade gracefully and reconcile asynchronously
  • Local Decisioning — cached models/policies for low-latency classification and gating
  • Identity Interfaces — support for biometric and device-based authentication where required
  • Hardware Flexibility — deployment on standard Linux edge stacks through to enterprise edge servers

Observability & Governance

Agentic systems require observability that can answer "what happened" with precision — including the coupling of decision → mandate → policy → action → outcome.

When an agent acts on behalf of a principal, the system must be able to prove:

  1. What was decided — the action selected and the alternatives considered
  2. Why it was allowed — policy evaluation, constraints, and approvals
  3. Who authorised it — mandate and delegation chain back to the principal
  4. What it caused — downstream operational and economic effects (including settlement)

Observability Stack

Traces

Distributed tracing extended for agent workflows, tool calls, and policy evaluation spans.

Metrics

Service and workflow SLOs: latency, success rate, exception patterns, routing outcomes, and settlement integrity.

Logs

Structured logs with correlation IDs linking intent, product state, payment intent, mandates, and execution outcomes.

Governance Controls

  • Policy Engine — declarative constraints for agent behaviour (spend limits, allowed actions, allowed partners)
  • Approval Workflows — human-in-the-loop for high-value, novel, or regulated actions
  • Evidence & Audit Export — compliance-ready exports linking mandates, state transitions, and settlement events
  • Kill Switch & Revocation — immediate authority revocation at user, org, agent, or system level

Regulatory Considerations

Agentic commerce operates in an evolving regulatory environment. The architecture is conservative by default: explicit authorisation, deterministic policy gates, and full auditability are required for machine-initiated execution.


Sequence: Agent-Initiated Purchase

The following sequence illustrates a typical agent-initiated purchase flow — showing how intent becomes an offer, how an agent obtains authority, and how settlement links to outcomes.

Figure 3: Sequence diagram showing payment intent, mandate evaluation, authorisation, and settlement

Explore the Technical Deep-Dive

The full specification expands on contracts, data models, and deployment patterns for partners and integrators.