The PlaybookAI Strategy

OpenCommerce Wired 5,026 Stores for Agents. Amazon included.

Three weeks after Amazon was reported to be blocking agents, a startup shipped one endpoint that buys from 5,026 stores. Amazon tests at 100%.

John TanOctober 30, 20265 min read

On October 30 a company called Agentcard shipped opencommerce, which it describes as the OpenRouter for commerce: one endpoint, every online store. The counter on its own site reads 5,026 stores.

Amazon is one of them.

Three weeks earlier, the question on the table was whether the largest retailer on the internet would serve agent traffic or block it. Shopify, Stripe and Walmart had publicly committed to serving it. Amazon was absent from the list, and Paul Graham called that absence the biggest opening he had seen in thirty years.

Somebody took the opening in twenty two days. Not by building an Amazon competitor. By routing through Amazon.

5,026

Stores reachable from one endpoint, per the vendor's own counter.

1%

Of every completed order, charged to the agent developer. Not to the store.

100%

Amazon's success rate on the public monitor, 45 days running.

The product is six lines of TypeScript. You hand your agent a buy tool, and the prompt does the rest.

A TypeScript code sample importing AgentcardClient and buyTool from the agent-cards commerce SDK, passing buy as a tool to generateText with the model claude-opus-5-5 and the prompt two packs of AA batteries from Amazon.
The SDK · 2 Nov 2026 Read the last line. The prompt names Amazon in plain English and the buy tool handles the rest.

From opencommerce.sh. Six lines, one of which is a sentence a person would say out loud.

The part that makes it checkable

Most agent-commerce launches are a video. This one ships a public monitor. They run the whole purchase flow at each tracked store, search to cart to checkout, and publish the pass rate and the response time next to the merchant's name.

A monitoring table listing DoorDash, Amazon, Walmart, Target, Best Buy and The Home Depot, each at a 100.0 percent success rate, with first response times between 4.0 and 16.2 seconds and 45 days of passing test history.
Live monitor · 2 Nov 2026 Amazon: 100.0%, last run 2 hours ago, first response 7.1 seconds, 45 days of green. Walmart answers in 4.0.

The vendor's own tests, published by the vendor. That is weaker than an audit and far stronger than a demo reel.

Read the Amazon row again. Not a workaround that sometimes clears. A tested purchase path, running every couple of hours, passing for a month and a half.

Who pays, and who never finds out

Figure 1. The money path
The store is in the network. Nobody asked it.
WHO PAYS THE TOLL WHO DRIVES CHECKOUT WHO GETS THE ORDER The agent OpenCommerce The merchant 1% of every order. One integration, not five thousand. Searches, carts, pays at the store's own site, with the saved card. Full price. No fee. No API to expose. No idea it happened. The fee is charged to the agent, not the merchant. A store can be inside this network without ever agreeing to be.
Nothing in the launch says Amazon has agreed to any of this. No Amazon statement exists, which is the same gap that made Graham's blocking claim an assertion rather than a report. Figure by nativefirst.
An Agentcard Vault card marked Approved showing a saved card ending 4242 and a line reading Checkout, Amazon with a completed checkmark, above a panel headed Complete purchases reading integrate with the wallet you are already using, more coming soon, with the agentcard, Link and Shop logos.
The wallet · 2 Nov 2026 A saved card, one approval, and a completed Amazon checkout. Link and Shop Pay are greyed out, listed as coming soon.
The Standard pricing plan card reading 1% of each completed order, for agents buying for real across every store and payment provider, with platform fees 1%, merchants every merchant, PSPs every PSP and API access no.
The toll · 2 Nov 2026 1% of each completed order. Merchants: every merchant. Nobody on that side of the line is quoted a price, because nobody on that side is charged.

The two halves of the business model. One takes the payment, the other takes the cut.

That is the structural point, and it changes the decision a seller is making.

A month ago the choice looked like serve agents or block them. It was not. The real choice is serve them yourself or get resold. Blocking the front door does not keep agents out of a catalogue that an intermediary can drive with a browser, a saved card and a test suite. It only means somebody else owns the integration, sets the price of access, and keeps the data about what agents were trying to buy.

Olivia Moore at a16z made the general version of the point the same week, about a different partnership. Her carve-out is the interesting part.

Olivia Moore
Olivia
Moore

"The dominoes are starting to fall on legacy platforms embracing agent traffic. With the exception of Amazon and their massive physical logistics as a moat, I think ultimately every platform is going to have to move this direction."

Olivia Moore  ·  X  ·  October 30, 2026

She wrote that the same day opencommerce launched with Amazon in the table at a 100% pass rate. Both things can be true. Logistics is a genuine moat and it is not the same asset as control of the buying surface. Amazon can own the warehouse and still lose the decision about which product the agent adds to the cart.

Why twenty seconds is the real spec

The speed claim is the part that separates this from the long tail of agent shopping demos. A cold route takes two minutes forty. A warmed one takes nine seconds.

Figure 2. Cold route versus warmed route
Two minutes is a demo. Nine seconds is a product.
COLD ROUTE STATED CEILING WARMED ROUTE 2m 40s 20s 9s 18x faster once the route is warm. Same stores, same carts.
Both figures are the vendor's, from its own feature page. Nine seconds is roughly how long a person waits before deciding a thing is broken, which is why this number and not the store count is the product. Figure by nativefirst.

What is still unproven

Four things, and an operator should hold all of them.

The tests are the vendor's. A 100% pass rate published by the company whose revenue depends on it is a claim, not an audit. It is checkable, which is more than most launches offer, and it is still self-reported.

The mechanism is not documented. Nothing on the site says how a purchase at a store with no agent API is actually completed. If the answer is driving the storefront, then every one of those 5,026 integrations is one layout change away from failing, and a store that wants it to stop can make it stop.

Coverage is uneven. On the October 30 snapshot of the monitor, eBay, McMaster-Carr and TikTok Shop were each labelled "no product search", Gymshark was "planned tests only" and Locale was "subscriptions only". A store in the count is not the same as a store an agent can shop.

Nobody has priced the fight. A 1% toll on somebody else's catalogue works until the catalogue's owner objects. There is no Amazon statement here, in either direction.

What to do about it this quarter

If you sell anything online, the question is no longer whether to support agent checkout. It is whether you would rather be a row in somebody else's monitoring table.

Two cheap moves. First, find out whether agents are already completing purchases on your site, which means looking at your own traffic for sessions that search, cart and check out in nine seconds. Second, decide what you want to happen when they do, because the default is now being decided for you by a third party with a test suite and a price list.

Blocking was never the wall. It was a toll booth, and somebody else is collecting.

Portrait of John Tan
John Tan ↗

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources
1opencommerce (opencommerce.sh), a product of Agentcard, Inc., announced 30 October 2026 by @keyserfaty on X: "a single endpoint that integrates 5k stores, including Doordash, Amazon and most shopify stores, with payments using Agentcard (Stripe Link and Shop coming soon)." Store count, the 1% Standard plan fee, the cold route 2m 40s versus warmed route 9s figures and the "under 20s" claim are all from the vendor's own site, captured 2 November 2026.
2monitoring.opencommerce.sh, the vendor's public merchant monitor, captured 2 November 2026: DoorDash, Amazon, Walmart, Target, Best Buy and The Home Depot each at a 100.0% success rate over a 45-day test history, with first response times of 16.2, 7.1, 4.0, 4.5, 7.8 and 14.6 seconds. These are the vendor's own tests of its own purchase flow, not an independent audit.
3The 30 October 2026 snapshot of the same monitor listed eBay, McMaster-Carr and TikTok Shop as "no product search", Gymshark as "planned tests only" and Locale as "subscriptions only", across 92 pages of merchants.
4Olivia Moore (a16z) on X, 30 October 2026: "The dominoes are starting to fall on legacy platforms embracing agent traffic... With the exception of Amazon and their massive physical logistics as a moat, I think ultimately every platform is going to have to move this direction." Written in reply to Instinct's OpenTable partnership, not about opencommerce.
5Paul Graham on X, 8 October 2026, on Amazon banning agents. As recorded in the earlier post on this question, Graham's characterisation remains the only source for Amazon's policy. No Amazon statement exists, in either direction, and nothing in the opencommerce launch indicates Amazon has agreed to anything.

Read next

AI is moving really fast. Keep up.

Read the Playbook