The Future of Shopping | Aidō Lighthouse

How to read the stack

The acronyms are not the point. The useful question is simpler: what does an agent need to do, in order, before it can buy or act for a customer?

Discovery Find the right agent card, MCP server, product catalogue, registry, API, or skill.

Interaction Call tools, ask another agent to do work, or use a browser session safely.

Product data Match the exact product, variant, GTIN, price, stock, and terms.

Identity Prove the agent is legitimate and acting under a real user mandate.

Payment Complete the purchase without exposing raw credentials or relying on a human checkout flow.

ARD belongs at the front. Before an agent can use MCP, A2A, WebMCP, APP, or a checkout protocol, it has to know those resources exist.

Discovery: ARD's job

Agentic Resource Discovery, announced by Google on 17 June 2026, is an open specification for publishing and finding AI capabilities across the web. Put simply, it gives agents a standard place to look.

A site publishes an ARD catalogue at /.well-known/ai-catalog.json. That catalogue can point to MCP server cards, A2A agent cards, AI skills, OpenAPI tools, registries, or other catalogues. Registries can then crawl and index those catalogues, so agents can search by intent rather than by guesswork.

Without a discovery layer, the market fragments quickly. One site advertises an MCP server. Another has an A2A card. A third has a product feed. A fourth has a registry. ARD is the page at the front of the binder: here is what we expose, here is where it lives, and here is the trust metadata to check before connecting.

Agentic Resource Discovery (ARD)

Discovery layer

Think of it as a catalogue, not a shopping or payment protocol. It does not replace MCP, A2A, WebMCP, APP, UCP, or ACP. It makes them findable.

What to publish /.well-known/ai-catalog.json with entries for agent cards, tools, skills, registries, and catalogues.

What to include Identifiers, display names, resource type, URL or inline data, representative queries, and trust metadata.

How agents find it Well-known URL first, then signals such as robots.txt Agentmap, HTML rel="ai-catalog", and DNS TXT records.

Why retailers should care It gives AI systems a controlled route to your machine-readable resources, instead of making them infer everything from page markup.

The protocol map

The names will change. The roles are steadier. This is the cleaner way to read the market.

ARD

Google + partners

Discovery catalogue for agentic resources: MCP servers, A2A agents, skills, registries, OpenAPI tools, and nested catalogues.

Draft · Jun 2026

MCP

Anthropic / open governance

A tool and resource protocol. Useful when an agent needs to call a server-side capability in a standard way.

Active

WebMCP

W3C Community Group

Browser-side tools exposed from the page, using the user's existing session. Powerful, but security needs careful thought.

Draft / trial

A2A

Google + Linux Foundation

Agent-to-agent delegation. A specialist agent publishes a card and another agent can hand work to it.

Stable

APP / AOCF

Klarna / open feed work

Product discovery and catalogue readiness: exact products, variants, prices, stock, GTINs, and purchase terms.

Emerging

TAP / Verifiable Intent / KYA

Visa, Mastercard, KYA networks

Identity and mandate: prove the agent is legitimate, the user authorised it, and the spend is within scope.

Active / emerging

UCP

Open industry standard

Checkout and discovery for agents: manifest at /.well-known/ucp, REST, MCP, A2A, and embedded transports. v2026-04-08 added multi-item cart sessions and standardised catalogue access endpoints.

Active · v2026-04-08

ACP

OpenAI + Stripe

Checkout sessions and delegated payments for ChatGPT shopping flows. v2026-04-17 expanded Shared Payment Token support to Mastercard Agent Pay, Visa IC, Affirm, and Klarna.

Beta · v2026-04-17

MPP / x402

HTTP-native payment layer

Machine payments over HTTP: MPP (Stripe + Tempo draft) uses HTTP 402 with a payment challenge header; x402 handles stablecoin micropayments with native Cloudflare edge support.

Emerging

VGS / Skyfire / PayOS / nekuda

Payment infrastructure

Tokenisation, agent wallets, request signing, PCI scope reduction, and merchant or agent-side payment enablement.

Live / emerging

A normal agent journey

A well-built shopping agent should not browse like a human and hope for the best. The better path looks more like this:

Find the resources Fetch the ARD catalogue, read the available agent cards, MCP servers, product feeds, and registries.

Choose the right interface Use MCP for server-side tools, WebMCP for browser-session actions, A2A for delegation, or a product feed for catalogue search.

Verify identity and trust Check trust manifests, agent identity, signed requests, and payment mandates before any high-risk action.

Complete the action Reserve stock, build a cart, request approval if needed, and complete payment using tokenised credentials.

The important point is that discovery comes first. Without it, everything else depends on guesswork, search snippets, or hard-coded integrations.

What retailers should do now

No retailer needs to implement every acronym on this page. Start with the basics, then add protocol support where it genuinely helps.

Make the public site machine-readable Clean Schema.org (v30.0), accessible product pages, useful robots.txt, sitemaps, and stable product identifiers.

Publish an ARD catalogue List the resources agents should know about: product feeds, APIs, MCP servers, A2A cards, skills, and registries.

Fix product identity Expose GTINs, SKUs, MPNs, brand, variant IDs, stock, price, delivery terms, and return constraints consistently.

Decide which actions agents may take Search is low risk. Add-to-cart, checkout, returns, and account actions need stronger identity, consent, and audit controls.

Review bot protection CAPTCHAs and WAFs often block legitimate agents before any protocol gets a chance to work.