Explainer · agentic commerce
Mastercard Agent Connect is a one-to-many commerce layer for AI agents, merchants and payment providers.
Mastercard introduced Agent Connect on 9 September 2026 as a way for participating merchants and AI agents to connect through one integration across product discovery, cart orchestration and payment enablement. The important architectural point is that Agent Connect is not itself the whole payment stack: Mastercard separates commerce connectivity from Agent Pay, Verifiable Intent and the merchant systems that remain authoritative for price, stock and fulfillment.
Published 12 September 2026 · launch announced 9 September 2026 · primary sources below
1 integration
Mastercard's stated alternative to many point-to-point AI-commerce connections
3 stages
Catalog discovery, cart orchestration and payment enablement
Protocol-agnostic
Mastercard says the architecture is intended to support multiple protocols over time
What Agent Connect is trying to solve
AI shopping breaks when catalog truth, cart truth and payment truth live in separate human-only interfaces
A shopping agent can recommend a product from stale page text and still fail at checkout because the current size is unavailable, the price changed, shipping is restricted or a required option was never selected. Agentic commerce therefore needs more than search relevance: the agent needs live, structured access to the merchant's source of truth.
Mastercard says Agent Connect can expose merchant-authorized product information, including dynamic catalog data, pricing and availability, then support cart construction with taxes, shipping and fulfillment details before the purchase reaches payment.
That turns discoverability into an execution problem. The winning merchant is not simply the page an AI model can mention; it is the merchant whose current commercial state can be read, checked and acted on without inventing missing details.
The commerce path
Six states an agent-ready merchant should make explicit
Discover
The agent queries merchant-authorized product information, including current catalog data and availability.
Recommend
The agent compares eligible products and proposes items that fit the shopper's stated intent.
Build cart
The commerce layer confirms item choices plus final pricing, taxes, shipping and fulfillment details.
Authorize
The buyer's payment and intent controls decide whether the proposed purchase is permitted.
Pay
The transaction completes through a supported secure payment credential and participating payment provider.
Reconcile
The merchant and agent retain enough structured order and authorization evidence to support confirmation, support and disputes.
Agent Connect vs Agent Pay vs MCP vs x402
Four layers that can participate in the same autonomous purchase
| Dimension | Agent Connect | Mastercard Agent Pay | MCP | x402 |
|---|---|---|---|---|
| Primary job | Commerce connectivity and orchestration | Payment authorization and execution | Tool discovery and invocation | HTTP resource payment |
| Typical data | Catalog, price, stock, cart, tax, shipping, fulfillment | Credentials, intent, authorization, transaction evidence | Tool schemas, arguments, results | Payment terms, asset, network, amount, settlement proof |
| Buyer journey | Discovery → cart → payment handoff | Authorize → pay | Discover tool → call tool | 402 challenge → pay → retry resource |
| What it does not replace | Merchant source-of-truth systems or payment controls | Catalog discovery and cart orchestration | A commerce network or payment rail | Identity, catalog semantics or full procurement policy |
Mastercard describes Agent Connect as protocol-agnostic. That is useful framing: commerce discovery, tool invocation, user intent and settlement are separate jobs. An agent might use MCP to call a merchant tool, Agent Connect or another commerce network to reach merchant catalog/cart state, and a separate payment layer to authorize and settle the purchase.
Why Anthropic's commerce blueprint matters here
The application layer and the commerce-network layer are converging without becoming the same thing
Anthropic released its commerce-agent blueprint on 2 September 2026 with reference shopping and merchant agents for retail, travel, telecom and ticketing. The shopping agent includes integration points for catalog, cart, checkout, customer preferences and order history, while payment is deliberately left to the merchant's existing checkout or an agentic-payments provider.
Mastercard says it plans to make Anthropic's commerce-agent blueprint available to merchants through Agent Suite for Merchants. That pairing is instructive: the agent application can reason over shopper intent and merchant tools, while a commerce connectivity layer exposes current product/cart state and a separate payment layer enforces authorization.
What merchants should expose to agents
Authoritative product identity
Stable product and variant identifiers, options and machine-readable attributes so agents do not invent equivalent SKUs.
Current commercial state
Price, stock, availability windows, taxes, shipping and fulfillment data sourced from the merchant system that will actually accept the order.
Cart rules
Required options, quantity constraints, bundle rules, geography restrictions and the exact state needed before checkout.
Authorization boundary
A clear transition between recommendation and consequential purchase, with user intent or organizational spending policy enforced outside the model.
Outcome evidence
Order confirmation, payment status, fulfillment state and error semantics that let the agent reconcile rather than blindly retry.
Provider-owned contracts
Use structured APIs, tool schemas and machine-readable policy instead of asking an agent to infer transaction logic from presentation HTML.
Primary sources
- Mastercard — Agent Connect product overview
- Mastercard — Agent Connect launch and Agent Suite for Merchants expansion, 9 September 2026
- Mastercard — Verifiable Intent and Agent Pay trust layer
- Anthropic — Building commerce agents with Claude, 2 September 2026
- Anthropic — commerce-agents reference implementation