On September 2 Anthropic published free, working code for building a shopping agent. Two days before that, the standard those agents publish into changed shape for the fourth time this year.
The build kit arrived on September 2. Anthropic released Claude Commerce Agents under an Apache-2.0 licence: two agent patterns, a shopping agent that searches a catalog and assembles a cart, and a merchant agent that handles inventory, pricing and marketing. It ships with reference builds for retail, travel, telecom and ticketing, and runs on the Claude API, Amazon Bedrock, Microsoft Foundry or Google Vertex AI. Shopify and Priceline are already running live agents on it. Visa, Mastercard and Accenture are named partners.
If you have two competent engineers, you can stand up a shopping agent this quarter and pay nobody for the code. That is worth saying plainly, because a lot of people selling into this market would rather you did not hear it.
Read what the release leaves out and the shape of the market gets clearer. No payment protocol. No checkout. No advertising layer. Those omissions are deliberate, and all three stay inside the retailer's own infrastructure, where the merchant relationship already lives.
The code stopped being the expensive part. Keeping it correct started being the expensive part.
Three of three model vendors now stop at the shortlist and leave the sale to you.
OpenAI retired Instant Checkout in March after five months and pushed ChatGPT back toward discovery, with retailers building their own surfaces. Google's AI Mode books hotels, and merchant-of-record status sits with the booking partner rather than with Google. Anthropic will not touch the wallet at all. That is a settled position now rather than a trend somebody is guessing at: the agent finds, the merchant closes, and the closing happens on a surface the merchant owns.
For anyone running a business, that is the good half of the news. The valuable end of the transaction stays with you, along with the customer record, the payment relationship and the ability to fix things when something goes wrong at 11pm. One condition attaches. The agent has to be able to find you and call you, which means you need a surface it can read.
Last time the route ran through ten names. This week the largest hotel brand in the world demonstrated the other route.
Our last briefing covered Google turning on hotel booking inside AI Mode with exactly ten partners able to complete the sale, and no application process by which an eleventh gets added. That was a story about a closed list.
On August 31 Hilton showed what the open route looks like. It put an AI Planner on its own site and app, joined Google AI Mode with Honors points redemption, shipped a ChatGPT plugin, and has a Claude Connector in development. In every one of those cases the booking redirects to Hilton.com. Published into three assistant surfaces, kept its own checkout, one week after the ten-partner list came out.
The second change is the one that costs money. On August 31 the Universal Commerce Protocol shipped its fourth spec release in a year, its first since April, carrying breaking schema changes. It added grocery features and laid groundwork for food and lodging: hotel discovery, room selection with rate options, booking with guest details, and payment scheduling. Shared components including token binding were redesigned. Anyone who implemented UCP by hand now has a migration to run.
Line the two dates up. Free, working agent code on September 2. A standard that moved underneath it on August 31. The free code does not move with the spec, and neither does the person who wrote your version of it.
Three vendors, three separate decisions, one answer.
Sources: OpenAI's retreat from Instant Checkout to discovery (March 2026, confirmed by Forrester in April); Google AI Mode agentic hotel booking with ten merchant-of-record partners (August 27, 2026); Anthropic's Claude Commerce Agents release under Apache-2.0 (September 2, 2026), reported by PYMNTS and MarkTechPost.
The build is a sprint. The maintenance is a job somebody has to hold.
Start with the part that argues against urgency, because your engineers will raise it and they are right to. Mastercard's August Signals report puts the share of consumers willing to let an agent complete a purchase autonomously at 10%, and finds 95% expressing some concern about agentic commerce. NMI's survey of 1,000 US adults, published a week earlier, put the ever-completed-a-purchase figure at 11%. Two independent reads landing a point apart is about as solid as consumer data on this gets. And on September 6 ABC News reported Australian retail specialists putting agentic shopping at least three years from standard practice there, with grocery arriving last rather than first, because substitutions are personal and one wrong item costs trust in the whole basket. Nobody should sell you this as revenue next quarter.
The other half of the same report carries McKinsey's figure of $3 trillion to $5 trillion in consumer spending mediated by agents by 2030, and adds that potential business-to-business agentic spending could run up to three times the consumer number. Anthropic says retailers running its commerce agents see 35% larger carts and 60% higher purchase completion. Those last two are Anthropic's own numbers with no independent benchmark behind them, so hold them as directional. The direction is consistent across everyone measuring.
Willingness to let an agent spend your money is small, well measured, and not the number that decides whether to act.
Sources: Mastercard Signals report "Encoding Trust", August 2026, for the 10% and 95% figures and for McKinsey's $3 trillion to $5 trillion of agent-mediated consumer spending by 2030; NMI survey of 1,000 US adults, published August 26, 2026, for the 11%. Both are published by companies that sell payment infrastructure, so treat them as vendor research. They were run separately and landed within a point of each other.
The case for acting is not the purchase number. It is the discovery number, and it is the cost of being late to a surface that takes months to publish properly. Which brings us back to August 31.
Four releases in a year, and the most recent one redesigned shared components rather than adding at the edges. If you build the agent surface yourself, the question worth settling before you start is who owns the next migration, and what happens to it when a breaking release lands mid-quarter alongside everything your engineers already committed to. Nobody buys maintenance for a problem they have not felt yet, which is why this week is useful. The problem now has a date on it.
There is a second cost, and it is uglier. On August 31 Rein Security published the full technical detail of a five-stage compromise of a major US retailer's AI shopping assistant. Indirect prompt injection through the product comparison feature. A search field that reached the backend without the intent classification applied to the chat box. Disclosure of the internal prompt and the tools the agent could call. Remote code execution in the agent's backend. Unrestricted Google Maps API keys lifted from app traffic. No privileged access was required at any point, and the whole chain ran through the same interface an ordinary shopper uses. The retailer was notified in March and had not fixed it by July.
Look at the first stage again, because it is the one that generalises. One path reached the backend without the checks applied to another path. Watching what an agent says is a different discipline from controlling what an agent can do. Worth knowing which operations your assistant is allowed to call today, and whether anyone has ever written that list down.
One more thing for the calendar. NIQ and Similarweb are building agentic commerce measurement across ChatGPT, Gemini, Google AI Mode, Perplexity and Claude, live in the fourth quarter of this year, connecting AI-driven discovery to traffic, conversion and verified sales. Today you cannot tell whether an AI platform mentioned you, sent you traffic or closed a sale. From Q4 you will be able to, and so will your board and anyone else who buys the data.
Publishing the surface correctly, and keeping it correct, is the part that does not come in a repository.
BeaconSpec is the discovery layer for agentic commerce. Point us at the REST, OpenAPI or GraphQL APIs your booking engine, catalog or ordering system already runs on. We turn them into one curated server an agent can call, and publish it as a UCP server card at your own /.well-known, so assistants can discover it without anyone at those companies picking you first.
If your team wants to build on Anthropic's blueprint instead, that is a reasonable call and the licence is generous. Be clear about what arrives with it. You get an agent. You do not get a published discovery card, a curated view of your catalog, a signed-in customer, or a plan for the next schema change.
Curation is the first piece we add. You decide operation by operation what an agent may ever see or do, and anything outside that list does not exist as far as the agent is concerned. Handing an agent your whole API is exposure rather than a strategy, and the Rein writeup is a fair description of where that ends.
Sign-in is the second. Real commerce needs the customer actually logged in before "move my reservation" or "place my order" means anything at all. We handle the standard login handshake, so the agent acts inside your customer's own account, with no shared credentials and no separate build for each AI vendor.
Run it hosted by us, or take a sealed container and run the whole thing inside your own walls if the data cannot leave the building. And because the specs move every quarter, keeping that surface current is our job rather than a ticket that sits in your engineers' backlog until it breaks.
A smaller, separate thing worth doing
Before an agent can shortlist you, it has to be allowed to reach you at all, and that permission sits in a text file most people have never opened. On September 15 Cloudflare's default flips for sites that never wrote their own rules. Answer three plain questions and our AI Visibility Checker writes a correct robots.txt for you.
Check your AI visibility →No sign-up. Copy the result and paste it into your site.
The agent became free the same week the standard broke.
Anthropic gave the market working commerce-agent code on September 2 under a permissive licence. Two days earlier the protocol those agents publish into shipped breaking schema changes, for the fourth time in a year. Both facts are true and they point the same way: the build is getting cheap, and the correctness is not.
What you own is the surface: a curated, standards-compliant interface any agent can find, read and buy through, with you deciding exactly what it may see and do, running wherever you need it to run, and kept current as the specs keep moving.
Three things are worth watching between now and December. Whether Google opens hotel booking past its ten partners or publishes a path an independent operator can implement directly. Whether anyone offers a hosted, maintained version of Anthropic's blueprint, which would change this calculation. And what EMVCo asks of merchants when the comment period on its draft agentic payments framework closes on September 30, since the centrepiece is an intent layer where consumer-authorised intent gets registered and updated before, during and after a transaction, with a Know Your Agent check under exploration alongside it.
BeaconSpec exists for exactly this shift: making your business discoverable and transactable by AI agents over UCP, on an interface you publish and control, kept correct as the standards underneath it keep changing.
See what we do • Try the free AI Visibility Checker • Read more articles
David Soden writes about automation, web strategy, and building durable technical systems for businesses. Photography via Pexels (Tima Miroshnichenko, Marcial Comeron, Sonny Vermeer). Figures cited are drawn from Anthropic's Claude Commerce Agents release under Apache-2.0 (September 2, 2026), reported by PYMNTS and MarkTechPost; the Universal Commerce Protocol's fourth spec release with breaking schema changes (August 31, 2026), from the protocol's own release notes and Search Engine Journal; Hilton's multi-LLM integrations and its redirect-to-Hilton.com booking policy (August 31, 2026), from Hilton's newsroom and Hospitality Technology; Mastercard's Signals report "Encoding Trust" (August 2026); NMI's US consumer survey reported by The Paypers and Retail Dive (August 26, 2026); Rein Security's published exploit chain against a US retailer's AI shopping assistant (August 31, 2026); NIQ and Similarweb's agentic commerce measurement announcement; EMVCo's draft framework for card-based agentic payments (September 1, 2026); and ABC News on the Australian retail timeline (September 6, 2026).