Essay · Agent-native infrastructure

Commerce, Rebuilt for Agents

Why consumer price comparison is a browser-era artifact — and what an agent-native commerce layer actually looks like.

April 14, 2026 Ge Jiaqi ~9 min read

For fifteen years, e-commerce has been optimized for a human with a cursor. The next fifteen years will be agents with intent. This is not a UX upgrade. It is a reshaping of the infrastructure underneath every transaction.

This post describes how we at xurprise think about the shape of that infrastructure, why the dominant framing most people reach for (cross-platform price comparison) is wrong, and where we think the industry actually has to go.

1. The comparison trap

When someone hears "products across multiple platforms," the usual reflex is price comparison. Show three sellers, highlight the cheapest, route the user there. It's what CamelCamelCamel does for Amazon, what Slickdeals does for deal hunters, what marketplace aggregators have done for twenty years.

It's also the wrong frame for an agent.

Cross-border price comparison hides a strong assumption: that the user can buy from whichever seller is cheapest. In 2026, that assumption is largely false. A shopper in Singapore can find a product on Taobao priced 30% below the Shopee SG equivalent, and still not be able to buy it — because the Taobao merchant doesn't dropship to Southeast Asia. The cheaper price is theater.

The same failure appears at every regional boundary: Lazada Philippines vs. Shopee Singapore, JD.com vs. Amazon US, Flipkart India vs. Rakuten Japan. E-commerce logistics did not globalize the way search did. Fulfillment is bound by geography, currency, duties, returns, and customer-service regimes that cross-border sellers rarely absorb.

For a human with a browser and time, you can let them compare and then redirect them to the version they can actually buy. For an agent, that is wasted tokens and wasted trust. An agent needs exactly one thing from a commerce layer: products that the user in front of it can actually purchase.

2. From consumer discovery to agent discovery

Zoom out one level. In the browser era, product discovery was a consumer task. People typed keywords into Google, clicked through blue links, landed on marketplace listings, compared specs, read reviews, and made a decision.

That stack was optimized for one actor (a consumer), one surface (a page), one mode (visual). Every design decision — image-heavy listings, above-the-fold hooks, trust signals shaped like social proof — was shaped around that actor.

Agents do not read that way. They query structured data, follow links programmatically, evaluate dozens of options in parallel, and make a choice without ever rendering an image. The surface they need isn't a listing — it is a feed.

The underlying shift is from consumer discovery to agent discovery. The question a merchant used to answer — "how do I rank on Google for my category?" — is being replaced by a new one: "how does my product become discoverable to Claude, ChatGPT, Perplexity, Cursor, and a long tail of custom agents that are already making shopping recommendations today?"

Two surfaces matter for agent discovery:

  1. Web surfaces — pages that crawlers can read, backed by schema.org/Product JSON-LD, llms.txt indexes, and stable URLs. Modern LLMs already treat these as first-class retrieval sources.
  2. Programmatic surfaces — REST endpoints and MCP (Model Context Protocol) servers that an agent can call with a shopping intent and receive structured results synchronously during a user's conversation.

Both surfaces need to be region-aware by default. An agent serving a Singapore user should see Shopee / Lazada / TikTok Shop SG. An agent serving a China user should see Taobao / JD / Pinduoduo. Not because we decided to silo the world — because that is what the user in front of the agent can actually buy.

3. The shape of an agent-native commerce layer

Built from first principles, an agent-native commerce layer looks like this:

That is what we mean by "agent-native." It isn't a visual redesign. It is a rebuilt stack where the primary reader is a program and the fulfillment reality is geographic.

4. Today: aggregating affiliate APIs, region by region

The first layer we are building is an aggregation layer that speaks to the official affiliate APIs of the major platforms in each region we serve.

Official affiliate APIs matter for three reasons.

First, they are sanctioned — the data they return is explicitly meant to be redistributed, which means legal compliance is inherited from the upstream API, not engineered at our layer. Second, they include attribution parameters — a click on our 302 redirect generates a commission event without any client-side gymnastics. Third, they have structured data — SKUs, prices, ratings, and specifications are already normalized inside each affiliate feed, saving us the expensive work of parsing marketplace HTML.

The trade-off is coverage. Official affiliate feeds tend to oversample commission-heavy products and undersample long-tail items. We accept this as an MVP limitation and compensate by treating the affiliate feed as one input — not the only input — to the discovery layer.

The geographic expansion order is deliberately not "North America first." With a small starting budget, validating an agent-native stack in less-saturated markets is an order of magnitude cheaper. We start where friction is lowest (China and Southeast Asia), build the discovery primitives, and expand outward.

5. Tomorrow: merchants publish direct

The interesting move isn't becoming a better aggregator of marketplaces. It is becoming the layer where merchants publish products directly for agent consumption, without routing through a marketplace at all.

Here is the observation. Amazon, Shopee, Taobao, JD — they own consumer-facing discovery because consumers have trained themselves to open those apps. Agents are a new class of buyer. There is no default destination. There is no "open Amazon first" habit. An agent searches across what is discoverable to it, and what it recommends becomes what gets purchased.

A merchant that publishes structured product data to the agent ecosystem early — before the ecosystem consolidates around one or two dominant agent hosts — will capture a disproportionate share of agent-driven transactions. The same merchants that had to pay Google and Meta for browser-era visibility will want a direct publishing path for the agent era.

The long-term direction for xurprise is to be that publishing layer. Merchants upload product catalogs with region-tagged fulfillment metadata. Agents discover them through our interface. Think of it as Amazon for the agent era, except decentralized by design: any merchant, any region, any agent.

This is a multi-year direction, not a quarterly deliverable. We are building toward it by first proving the agent-discovery stack works via affiliate aggregation, and by designing every data model today so that a direct-merchant publisher can plug in later without a redesign.

6. Why now

Three things had to be true for this bet to make sense. All three became true in 2026.

Agents reached consumer adoption. Claude Desktop, ChatGPT Apps, Cursor, Perplexity, Gemini — at least five mainstream agent hosts are in front of tens of millions of users, each of whom routinely asks an agent to find, compare, or buy something.

Structured data standards matured. schema.org/Product has been stable for years; llms.txt (Jeremy Howard, 2024) gave LLM crawlers a convention for site indexes; the Model Context Protocol (Anthropic, 2024) gave agents a standard way to call external tools with typed parameters. The plumbing an agent-native site needs is a few hundred lines of configuration, not a research program.

Affiliate networks opened up. Taobao Ke, JD Alliance, Duoduojinbao, Involve Asia, CJ, Awin, and Impact all expose official APIs that a small team can integrate against in weeks, not quarters. The legal and technical friction that made this work impossible in 2018 is gone.

The window is not permanent. Once one or two agents become default shopping interfaces (the way Google became the default search interface), marketplaces will rush to build their own direct agent APIs, and the first-mover aggregation play will compress. The bet is on the aggregation-to-direct-publish transition playing out over a few years, and on being the layer that holds through that transition.

7. What's live, and what's next

Live today:

Coming weeks:

Coming months:

If you are an agent author, a merchant, or an affiliate partner and you want to plug in early, we would like to hear from you.

E-commerce is about to be rebuilt. Not as a better search, not as a better app — but as a new layer between merchants and agents. xurprise is that layer.

— Ge Jiaqi, founder, xurprise · April 14, 2026