Anthropic Just Shipped the Commerce Agent Blueprint. For discoverGPT, That’s the Validation We’ve Been Waiting For.

This week Anthropic shipped a full reference blueprint for building commerce agents on Claude, and it’s now sitting on GitHub for anyone to pull down. Two agent types ship out of the box: a Shopping Agent that sits on a retailer’s own site or app to help customers find products, keep context across a conversation, and carry them through to checkout, and a Merchant Agent for the store’s own team to check performance, manage inventory alerts, and run pricing and promotions. Alongside them: four working vertical demos (retail, travel, telecom, entertainment), a Claude Code plugin that scaffolds a custom agent from the blueprint in days instead of months, and three ways to run it depending on how much infrastructure a team wants to own (the Messages API, the Claude Agent SDK, or Anthropic’s own Managed Agents). Shopify, Visa, Mastercard, and Accenture are already building on it, and Anthropic is reporting meaningfully higher purchase-completion rates for shoppers who use it.

That’s a lot shipped in one week. What I want to talk about is the part that matters most to us.

What Anthropic deliberately didn’t build

Read past the announcement and into the actual repo, and the design choice that stands out isn’t the agent logic. It’s what Anthropic chose not to own. The Shopping Agent doesn’t ship a search engine, a recommendation model, or a catalog format. It ships an interface: a deployment implements a StorefrontBackend over its own catalog, cart, order, and policy systems, and the agent’s skills call into that backend server-side. The model never touches your database directly. It reads whatever your backend hands back. A team can stub most of it and implement only search and product detail to get a working pilot running, and the harness doesn’t change a single prompt byte either way.

In other words: Anthropic built the agent. They left the door wide open for the discovery layer underneath it.

Why that’s the best news I’ve read all year

That’s precisely the layer discoverGPT already is. We didn’t build a chatbot that sits on top of commerce, we built the discovery engine underneath it: trimodal search, virtual try-on, catalog enrichment, and agentic feed generation, exposed as 40 MCP tools, built to be the backend a storefront (or an agent) calls into. Anthropic just published a reference architecture whose single biggest integration point is exactly the thing we’ve spent the last two years building. Any team that stands up a Shopping Agent from this blueprint can implement their StorefrontBackend search and product-detail methods by vibe-coding straight into discoverGPT’s MCP tools, and get trimodal search and virtual try-on on day one instead of building a catalog-search stack from scratch. That’s not a hypothetical integration we’d need to go pitch. It’s plug-and-play against an interface Anthropic just open-sourced.

That’s genuinely great news for the industry, independent of us. A credible, well-funded reference implementation, backed by Shopify, Visa, Mastercard, and Accenture, moves “an agent shops directly on the merchant’s own storefront” from thesis to default architecture almost overnight. It gives every merchant and every agency a fast, safe way to stand up a commerce agent instead of building bespoke agent infrastructure in-house. The whole category moves faster because the hardest, least differentiated part, the harness and the guardrails, is now solved and shared.

Why the catalog layer still has to be right

The blueprint also makes a quieter point, and it’s the one that validates why we built discoverGPT the way we did. Because the model only ever sees what the backend returns, the quality of that backend is the whole ballgame. An independent read of Anthropic’s own demo catalog found exactly this: a weight field recorded five different ways across products (254 g, 2.1 kg, 35 lb, 2 lb, 2 x 9 oz), so a shopper filtering for “700 lbs” of weight capacity got the filter silently dropped instead of matched, and 63 attributes that actually determine fit and suitability, like under-bed clearance or tent floor area, existed only in unstructured spec sheets, invisible to the agent entirely. That isn’t a knock on Anthropic’s demo; it’s the expected result of any catalog that wasn’t built for machine-precision search from the start, and it’s precisely the gap discoverGPT’s enrichment and trimodal search were built to close before an agent ever asks the question.

Put together, the blueprint and that finding say the same thing from two directions: the harness is now solved, and the discovery layer underneath it is the part that decides whether the agent actually works. That’s the layer we’ve been building the whole time.

Where this leaves us

Anthropic just handed the entire industry a fast, credible path to put a shopping agent on their storefront, and left the one integration point that matters most, search and discovery, open for whoever built it best. We think that’s us. Over the next few weeks we’ll be publishing a reference StorefrontBackend implementation on top of discoverGPT’s MCP tools, so any team building on Claude’s commerce blueprint can wire in trimodal search, virtual try-on, and enrichment without writing that layer themselves.