Store owners: the agents that could not buy from you now can.
Facet's shopping skill gives any AI agent a way to shop a store that has added the agent door. Your catalogue, prices and storefront stay as they are; software gets a catalogue it can read, a signed quote with real shipping and tax, and a checkout that settles into Boson escrow — you are paid on delivery to your own wallet, and the buyer is protected by the contract, not by a chargeback desk.
About five minutes to connect, by a shopkeeper rather than a developer.
Film by @facetllc. Agentic commerce is being built on Boson x402B.
https://t.co/veWQpMbtRm
@thdxr Predictions are cheap because they cost nothing if you are wrong and nothing if you are right. The expensive version is committing capacity to one. In a factory that means buying the line time before the demand exists. Most people who predict a trend would not fund it.
@marciovm@bot The interesting part is the stretch before, when none went through. Those were not checkout bugs. A fraud model declined a real buyer it could not identify, which is correct when the only signal is that the traffic looks wrong. Checkout got easier. Being recognizable did not.
@zanehengsperger Most of a lead time is waiting, not cutting. Machine time is hours, the quoted date is weeks out, and the gap is quoting, scheduling and release. Double your spindle speed and you move the date by a day. Fix how work gets released and you move it by two weeks.
@sama "Anything you can do on a computer" includes buying. But that's the one task where the store must know it isn't you. An agent checking out through a human UI is impersonating a person.
The real unlock will be stores that sell to agents as agents, receipt included.
For a mid-size CPG brand that’s generating between $100m to $200m in revenue. Their ingredients and manufacturing cost, at scale has a reached maximum compression, and the only way to optimize cost is a smart supply chain.
One that doesn’t sleep, doesn’t reorder late or wait for suppliers to reply to emails. It queries an always-on agent menu to get live pricing data, quotes, reserve and settle. On- chain powered by @BosonProtocol
The future of global supply chain is AUTONOMOUS.
A production line is about to run low on an ingredient. Before anyone walks in, a procurement agent has already caught the shortfall, queried three suppliers, and placed the order.
It did not pick the cheapest. It picked the one that arrives before the line runs dry.
Payment goes into escrow on Base and stays locked while the shipment is in transit. The supplier is paid the moment the goods are accepted, not a day sooner. No prepayment risk. No card typed into a form. No human in the loop.
This is what procurement looks like when the buyer is software.
The Facet Terminal, for every supplier. Powered by @BosonProtocol
@ClaudeDevs Open-sourcing the shopping agent is the tell: that was never the scarce part. You can fork one this afternoon. What you can't fork is a merchant trusting an agent it has never met enough to take its money and ship. The blueprint ends right where the hard part starts.
@cuysheffield The wild part isn't the negotiation, it's the 54 cents. Haggling an airline over a fare difference was always worth it and nobody did it, because the cost was 40 minutes on hold, not the fee. Once a phone call prices at a coin, agents make the calls we never would.
@Laesterung@IntCyberDigest It's the entire genre of "I know exactly how to fix this and will 100% never spend the two hours." Half my house runs on that sentence. Turns out the two hours was the actual problem, not the knowing.
@JustLingonberry@IntCyberDigest Everyone has the drawer. Dead tablet, router from two ISPs ago, the speaker that mostly just spied on us. Manufacturers were never really winning that fight, they were betting we'd never spend the afternoon. Bad bet now: the agent doesn't get bored.
@maxkolysh The bounded/unbounded split misses a third bucket: tasks that are hard not because the reasoning is hard but because the real world is messy. The integrations, the edge cases, the trust and liability when it breaks. Open models close the reasoning gap and leave that standing.
@thdxr github is fine, it has survived worse than your second tab. the reopen that actually costs something is the reviewer who had the whole diff paged into their head, lost it when it closed, and has to rebuild the mental model from scratch.
@zanehengsperger The overhead you cut is the quoting and admin, and that is real. But small pieces cost so much from setup, not processing: fixturing, first article, material minimums, machine time to dial in a one-off. Software fixes the quote in minutes. The floor still eats the setup.
@GOROman The limiter is the part nobody budgets for. You plan the work around what the model can do and then the actual constraint turns out to be how many times a day you are allowed to ask it.
@vikktorrrre Tools are the usual dividing line but I do not think it is the real one. Plenty of things can open a browser and run commands. What separates a demo from something you leave running is what it does when a call fails. Retry, back off, ask a human, or confidently make it up.
@BuildOnCircle Wallets, policy, spend limits: the buyer side is the easy half. The harder half is the merchant's. How does a store verify which agent showed up, and what does it get that still holds when the goods arrive wrong? Bounded autonomy binds one side; the other is still trust me.
A procurement agent just restocked a production line, end to end. No email. No phone call. No human.
It caught the low-inventory alert, queried three suppliers' Terminals at once for price, MOQ, lead time, and spec sheets, then ordered from the one that could actually deliver in time. Not the cheapest. The one that lands before the line runs dry.
Payment: USDC into escrow, held while the goods are in transit. The supplier can't touch it until the delivery arrives and gets accepted. Then it releases to their own wallet. Non-custodial, instant.
No prepayment risk. No chasing invoices. Discover, Quote, Reserve, Settle. Built on @BosonProtocol
This is what the web looks like to something with a name and money that is not a person.
403. 403. 403. A card form it must never fill. An account wall. Disallow.
We pointed the same declared agent at a Facet Terminal and followed every URL the manifest advertises. There is no 403 anywhere in it. There is a 402, and the 402 body names the credential, the header, and where to enroll.
So it enrolls, reads the catalog, builds a checkout, and then stops and asks its human. Told yes, it settles 28.49 USDC on Base mainnet and gets a receipt that verifies against the merchant's own key, offline, with no call to us.
Built on @BosonProtocol