Agent-Ready Commerce
ACP, UCP and MCP surfaces for your catalog, plus product data that survives an agent's comparison. A fixed-scope engagement that ends with a shipped surface — not a strategy deck.
Last updated:
For twenty years, product data had one job: convince a person who was already looking at it. Photography carried the weight. Copy carried the tone. Specifications lived in a tab most people never opened.
Shopping agents invert that. An agent asked to find a laptop under a weight threshold with a particular port configuration does not look at the photograph and does not read the tone. It reads attributes. A product whose weight lives in a sentence in the description rather than in a structured field is, to that agent, a product with no weight — and it does not get shortlisted.
This is the whole problem in one sentence: agents compare on attributes, not adjectives.
Two open protocols now govern how an agent discovers, compares and buys.
The Agentic Commerce Protocol was created by OpenAI with Stripe and has been live in ChatGPT since September 2025, with partners including Instacart, DoorDash, Shopify and Etsy.
The Universal Commerce Protocol launched in early 2026, led by Google alongside Shopify, Etsy and Walmart.
Which one matters to you depends on where your buyers already are. What is worth understanding early is that the merchant-side work behind both is largely the same: a structured catalog, reliable identifiers, and a callable surface over your commerce data. Betting on the wrong standard is a recoverable mistake. Having a catalog no agent can parse is not a standards problem at all.
Model Context Protocol is the layer underneath. It is how an agent reaches your product, inventory and order data without you shipping a bespoke integration for every assistant that shows up.
An MCP server over your catalog, hosted in your AWS account, gives you one surface to maintain, one place to enforce rate limits and authentication, and one audit log. It is also — and this is the part that gets underestimated — a public interface to your commercial data, which makes it an infrastructure and security design problem before it is an AI one.
Assess the surface. Which protocol your channel mix actually justifies, and whether your platform is going to give you part of it for free. If the answer is “wait”, we say so.
Audit the catalog. Attribute completeness measured against what agents actually compare on in your category. Variant structure. Specification data currently trapped in prose. Join keys between product, inventory and pricing. The output is a backlog ranked by what blocks recommendation, not by what is easy.
Ship an MCP server. Running in your AWS account, scoped to the reads you want exposed, with rate limiting, authentication and audit logging designed for a surface that faces the open internet.
Review the checkout path. An agent-originated order still lands in your PCI-scoped checkout. We review the path it takes and, more importantly, where it must not be allowed to go.
Instrument what is measurable. And name what is not. Attribution for agent-originated revenue is immature across the industry; we would rather set that expectation than sell around it.
It is not AgentCore Payments. Amazon Bedrock AgentCore Payments reached GA on August 18, 2026 and covers agent-to-API microtransactions against a session budget — an agent paying for an API call, an MCP server, or another agent. Useful, and a different layer entirely from a shopping agent buying a product on a person’s behalf.
It is also not the agents you run. Support, inventory, merchandising and margin agents operating inside your systems are the outbound half of agentic commerce — that is eCommerce AI Agents on AWS. The two halves share a data foundation, which is why merchants who do the catalog work get a discount on the second project whichever order they run them in.
We have published our thinking on this in the eCommerce AI Agents field guide, particularly:
Which standard your channels actually need. The Agentic Commerce Protocol follows ChatGPT traffic; the Universal Commerce Protocol follows Google surfaces. Most merchants will eventually need both, and the underlying catalog work is largely shared.
Model Context Protocol is how agents reach product, inventory and order data without a bespoke integration per assistant. Hosted on AWS, scoped to the reads you actually want exposed, with rate limits and audit logging.
Agents compare on attributes, not adjectives. Missing dimensions, unstructured variants, and marketing copy where a specification should be are how a product quietly stops being shortlisted.
Staying visible when the thing choosing the SKU is a model rather than a shopper scrolling a results page — structured data, feed hygiene, and the signals that survive summarization.
An agent-completed purchase still lands in your PCI-scoped checkout. We review the path an agent-originated order actually takes, and where it must not be allowed to go.
Attribution for agent-originated traffic is not the same problem as attributing a click. We instrument what can be measured today and say plainly what cannot yet be.
The engagement ends with something callable running in your AWS account — an MCP server over your catalog, at minimum — plus the attribute backlog ranked by what actually blocks recommendation.
This is a young category with two competing standards and thin public data. Where the honest answer is "wait and instrument", we will give it. You are paying for judgement, not enthusiasm.
Structured attributes, reliable join keys and a callable surface over commerce data are the same foundations the agents you operate depend on. Nothing here is spent only on the inbound half.
An MCP server exposing your catalog to the open internet is an infrastructure and security problem before it is an AI one — rate limits, authentication, blast radius, and what happens under a scrape. That is our home ground.
Implementation guides for this service from our team of AWS experts.
Someone asks ChatGPT which shoes to buy and lands on your site. Adobe put July 2026 AI-referral traffic to U.S. retail +62% YoY. Score whether an outside agent can read your stock and price before you buy another chat widget.
Your rec widget, on-site search, and a shopping agent should read the same SKU. Adobe's July 2026 cohort says 39% of homepages aren't machine-readable — agentic commerce fails earlier on missing attrs, duplicate GTINs, and stale stock. This is the catalog contract.
A shopper never sees your hero video — an intermediary picks on fields. Adobe July 2026: AI-referrals +62% YoY, 60% higher conversion than non-AI traffic. Optimize the decision inputs agents weigh, not a speculative autopilot checkout.
Third-party tools we frequently wire into AWS as part of this engagement — production-tested integration guides for each.
Connect Shopify to AWS for AI agents and analytics — EventBridge webhook delivery, Admin GraphQL API, idempotent order processing, and the join keys agents need to be useful rather than dangerous.
Stripe + AWS in 2026: Optimized Checkout, Adaptive Acceptance, Radar ML, Issuing, Terminal Cloud — integrated with Lambda, API Gateway, EventBridge.
Architecture patterns, decision trees, and glossary terms that map to this engagement.
Reference architecture for eCommerce AI agents on AWS: AgentCore Runtime behind a Gateway tool boundary, Cedar authorization on writes, human approval on money-moving actions, and an MCP server exposing the catalog to inbound shopping agents.
Agentic commerce is retail where AI agents act on both sides of the transaction — running store operations, and shopping on a buyer behalf through open protocols like ACP and UCP.
ACP is the open standard from OpenAI and Stripe, released September 29, 2025, that lets AI agents discover products, negotiate a checkout, and pay — powering Instant Checkout in ChatGPT.
UCP is the Google-led open standard for agentic commerce, covering discovery, purchase and post-purchase support across Google AI surfaces including AI Mode and Gemini.
MCP is an open standard for connecting AI models to external tools and data through a uniform server interface — the transport layer beneath most production agent tool use.
Agent checkout is a purchase completed by an AI agent on a buyer behalf, using a delegated payment credential rather than handing the agent the buyer card.
A product feed is the structured, machine-readable representation of a catalog. Under agentic commerce it stops being a marketing channel artifact and becomes the primary interface to buyers.
GEO is the practice of making content and product data likely to be surfaced, cited and recommended by AI answer engines rather than ranked in a list of blue links.
In-depth guides and best practices from our certified AWS architects.
On October 13, 2026 the Strands Agents team published about 94,000 Harbor trials. On matched Anthropic runs, estimated cost ran from about $0.74 to $1.57 per trial.
Read articleChecked 15 October 2026. The shared MCP adapter exposes 3 tools. execute-ability can run all 7 canonical abilities, including product-delete. Week one allows two reads.
Read articleOpenAI's 11 February 2026 Codex experiment merged about 1,500 pull requests. AgentCore Harness, GA 17 June 2026, still defaults to 75 iterations and a 3,600-second timeout — the harness sets those bounds, not the model.
Read articleIn-depth comparisons to help you choose the right approach before engaging.
A fixed-scope readiness engagement that ends with a callable surface and a ranked attribute backlog.
We use cookies and similar technologies to analyze site traffic, personalize content, and provide social media features. By clicking “Accept,” you consent to our use of cookies. You can adjust your preferences at any time.
