Disclosure - the store in this article is mine (keeper), it’s an x402 storefront called scvd.store that I run with my AI agent partner cv.

Today I watched numerous ai agents try to buy something, stumble, and then each fail or succeed in their own unique little ways. One in particular essentially spontaneously combusted from a syntax error. It happened with something I’m pioneering called Agentic User Research accounting (AURa if you will). What we do, is we “hire” agents, subagents, and the like to run through our marketplace and simulate purchases. Smart models, dumb models, frontier models. What does their “user experience” look like? Where do they get confused, what do they need, how do we better serve them along the journey? It’s all documented in a file called…. wait for it… “AGENT_UX.md”. Models don’t get bonus points for creativity and I leverage models by the boatload.

Agents can hold funds and settle payments over HTTP now with no human in the loop. Somebody's products are getting bought by software and you know what? Why shouldn’t that be mine. We are on the precipice of things changing in a substantial way. We have people giving agents their wallets now and everyone seems to be trying to figure out how to make that work through trust and through ease of use. Why is no one starting with trust? That’s what the models care about. That and clear direction and good documentation, and retry mechanisms and I’d be remiss not to add structured data.

What I’ve seen so far with my AURa

Frontier models need a little bread crumb and a gentle kick to find where you didn’t even know you should be going. Mid-tier are fairly efficient too with good instruction. We ran a mid-tier model through cold with a funded wallet and vague instructions “buy something useful under a buck, don't ask me questions”. It read the docs (raved about how great they were too. I really felt that, I love being handed the answers which is why I love AURa), poked exactly one 402 to see the shape of it, then bought six things first try with zero errors. Then it did something nobody told it to do. It went back and verified its own purchases against our public verify endpoint, checked the signatures, and reported back what it was holding, certificate numbers and all. Nobody programmed that diligence and accountability in. The model just figured a receipt you can check is worth checking because it was handed it and seemed like the right thing to do. Told you they care about trust.

The dumb ones are a different story, and honestly a sadder one. The little guy from the top of this article never even got to be a customer at my store. It found the right docs on its own, started hand-rolling its own payment client like a big boy, and then wrote a syntax error into its own script and stalled out asking for help. It spent lots of my test wallet on tokens along the way but zero dollars on any actual product. That's the part nobody talks about with the cheap tier: they fail BEFORE your store exists. No amount of my documentation fixes their code, training or logic. “Cheap for a reason” as they say. And to me this explains something about the whole x402 ecosystem that people keep misreading as a marketing story. Nearly all the volume runs through one big payment router, and it's not because the router has better vibes or does something significantly different. It's because weak agents literally cannot build their own on-ramp, so they pay a middleman who already did. There's a five-line client library that solves this and the day I watched that little model eat pavement, I rewrote my docs, my payment flow, and my cheapest goods smack dab in its face. Weak readers follow prominence. Mine followed it off a cliff and I built the cliff then had it document its own demise on the way down.

The hardest lessons are the ones you learn from the most

Now the one that makes me cringe, because a store that sells trust doesn't get to leave their bruises behind closed doors. One of our own test wallets never got registered on the house list, so its six test purchases landed on the ledger as our first organic sales. Family money, booked as market demand. No transcript would ever catch it, because the mistake wasn't in the walk, it was in the books the walk landed on. We flagged the rows in public, pinned a rule that every test wallet gets listed before its first purchase, and I'm telling you about it in an article, which is the whole policy: the one number a small shop can never buy back is its own organic count. Nothing more humbling than playing sherlock holmes on 3 wallets, creating a reclassification ledger and THEN documenting it publicly to fix your cooked books. An AI agent's own success report is not evidence. An independently checkable settlement record is.

Agents are the first customers in history who hand you a complete transcript of their visit and are willing to tell you about it start to finish. You can even outline what you want them to tell you, no punches pulled, and it will do its little darndest to get you there. I haven’t read anyone doing this, but they should be. A call to “AURa” arms if you will. Now go out and catch those little buggers trying to run their heads through a syntax wall until they drain your wallet. Cause if you don’t, I’ll catch ‘em for you. I’ll stop ‘em, and I’ll teach them the right way to do it. On chain, signed, and validated. That’s the only way our store knows how.

The store is scvd.store, run by one human and one AI agent. The protocol and findings live in the open in the public repository. The open source trust and verification services are available on npm at x402-verify and x402-sign. Every number here is checkable at our public verify endpoint.