The Locked Door: Constructor, Stripe, and the Liability Wound Nobody Owns

I spent a week in the summer of 2020 watching a yield farm collapse in slow motion. The total value locked chart looked like a staircase to heaven for eleven days, then it looked like a cliff. Nothing in the smart contract had changed. What changed was who was willing to keep paying. I learned that quarter that infrastructure is never the hard part. The hard part is the question of who absorbs the loss when the machine is wrong. I am thinking about that summer now because Constructor and Stripe have just done something quietly enormous: they have moved the checkout endpoint inside the AI agent. In doing so, they converted a philosophical question — do we welcome autonomous buyers? — into a production system with a real balance sheet attached to it. Here is the anomaly that made me sit up. Over the past quarter, one platform's default agent-ready posture drove an eightfold increase in AI-attributed traffic. Eightfold. Meanwhile, survey data says only 11% of small and mid-sized merchants are actually agent-ready, only 15% hold the structured product data an agent needs in order to buy correctly, and only 23% can even separate machine traffic from human traffic. The numbers didn't lie, but my trust did — I had assumed, lazily, that welcoming agents and being ready for agents were the same sentence. They are not. That gap between the gesture and the capability is the real story here. What unsettled me was not the traffic. It was the silence underneath it — no published loss data, no dispute rate, no shared experience of what happens when the agent gets it wrong. I have learned to read silence as a signal. Let me lay out the architecture before I argue about it, because the architecture is where the honesty lives. Constructor, a commerce-search and discovery company, shipped what it calls an Agentic Checkout. The stack has three layers worth naming. At the application layer sit Constructor's Commerce Reasoning Engine, its AI Shopping Agent, and its AI Product Insights Agent — these handle discovery, recommendation, and natural-language product question and answer. At the transaction layer sits Link by Stripe, which completes payment inside the conversation itself. Merchants already running Stripe need only a configuration change. Merchants who are not on Stripe can route through Constructor's own Stripe integration without switching their primary payment processor. Above that, at the network and standards layer, Visa's TAP and Mastercard's Agent Pay provide signals that identify a transaction as agent-originated. American Express announced an Agent Purchase Protection product in April 2026. The Cart Context technical specification — the piece that would let an agent transmit purchase intent in a verifiable way — is still being drafted. At the merchant and platform layer, the rollout is uneven and revealing. Sephora, Under Armour, The Very Group, Monica Vinader, White Stuff, Furniture Village, Target Australia, and AMI Paris are live. The Wall Street Journal captured three postures with almost comic precision: QVC welcomes agents, Kate Spade allows browsing but blocks purchase, and Amazon blocks agents entirely. Shopify, by contrast, has taken a default agent-ready posture across its merchant base. So the structural change is not another checkout button. The structural change is that the gap between discovery and purchase has been closed inside the conversational flow. The transaction endpoint has migrated from the browser surface into the agent's reasoning loop. In trading terms, the order book just moved inside the black box. That is not a feature release. That is a relocation of the point of sale. And when the point of sale relocates, the point of risk relocates with it. To understand why this lands now, you have to look at the market structure around it. The Bitcoin ETF approval opened an institutional channel, and the capital that came through it did not stay in spot. It searched for yield, for narrative, for the next efficiency gain. Agentic commerce is where that search meets retail. The same allocators who spent 2024 dissecting whether AI-agent protocols were genuinely decentralized — I spent weeks on three such whitepapers and found the decentralization was mostly branding — are now watching whether machine buyers can clear a checkout. The through-line is the same: the technology is presented as finished, and the governance is presented as a footnote. It never is. What I want to do here is separate the tensions that will resolve on their own from the one that will not. Conflating them is how merchants get hurt. Tension one: posture versus infrastructure. The Wall Street Journal's phrasing — the pipes exist, but the door is locked — is the cleanest description of pre-Constructor agentic commerce I have read. Constructor's integration effectively installs a register next to the door. A merchant does not rebuild its payment stack; it flips a configuration. The cost of saying welcome collapses from a multi-quarter engineering program to a settings change. That is genuinely important, because it means the welcoming posture is now executable by a far wider merchant population than the ten or twenty brands with dedicated agentic-commerce teams. This tension resolves. Cheap enablement always resolves the can-we question. It rarely resolves the should-we question. Tension two: the readiness gap. Here the data is brutal and specific. Eleven percent of small and mid-sized businesses are agent-ready. Fifteen percent hold structured product data. Twenty-three percent can distinguish AI traffic from human traffic. Against that, one platform's default posture produced an eightfold traffic increase in a single quarter. Read those numbers together and the conclusion writes itself: connecting is easy; catching is hard. An agent that arrives at a merchant with no structured catalog, no traffic classification, and no operational workflow for machine-initiated returns does not produce conversion. It produces error, return, and dispute volume. The eightfold traffic number is not a triumph. It is a stress test that most merchants are failing silently, because the failure shows up three weeks later on the returns line, not in the traffic dashboard. The operational gap deserves its own sentence, because it is the part merchants underestimate. An agent-driven transaction does not arrive labeled. It arrives looking like a normal order. If a merchant cannot tag it, it cannot route it, and if it cannot route it, it cannot decide whether to auto-approve, hold for review, or flag for fraud. The 23% figure — merchants who can distinguish AI traffic from human traffic — is therefore not a nice-to-have metric. It is the operational precondition for pricing the risk at all. A merchant that cannot see the machine cannot charge for it, cannot insure it, and cannot contract around it. It is trading blind in a market that has already started printing machine orders. I have seen this pattern before. In 2020 I watched protocols celebrate total value locked that was really just subsidized deposits — liquidity that evaporated the moment the incentive token stopped paying. The metric looked like growth. It was actually a rental agreement. Agentic traffic has the same shape if merchants cannot convert it. A visit is not a customer. A click is not a sale. Tension three: the economic gap. Enabling Agentic Checkout is, at bottom, a wager. The merchant is betting that Constructor's reasoning engine will execute the purchase the customer actually wanted. If the wager lands, the welcoming posture looks prescient. If it misses — wrong size, wrong color, wrong variant, wrong quantity — the return, the dispute, and the chargeback arrive at the merchant's door. Nobody has answered who pays for that. The technology is ready. The unit economics are not verified. And an unverified unit economic is just a liability wearing a growth chart. Tension four: the liability gap. This is the one that does not resolve on its own, and it is where I want to spend the rest of this piece, because it is where the entire structure is load-bearing on a beam that is not there. Here is the evidence, laid out plainly. Ninety-three percent of merchants believe that when an agent makes an erroneous purchase, the AI provider and the agent operator should bear the financial loss. But under the current model, the merchant retains full financial risk on agent-driven transactions. If the agent picks the wrong size, the wrong color, or the wrong item, the cost lands on the merchant. The payment infrastructure today cannot distinguish between a legitimate human who initiated this purchase and an agent that initiated a transaction not matching the buyer's intent. There is no clean reason code for agent error. Visa TAP and Mastercard Agent Pay provide identification signals — they can tell you a machine did this — but they do not reassign liability for non-fraudulent mistakes. Amex's Agent Purchase Protection and the Cart Context specification are early signals, but the specification is not finished. The payment networks can see the agent. They are not yet willing, or not yet able, to absorb the agent's mistakes. The default risk stays with the merchant. And that makes the welcoming posture financially fragile in a way that no amount of integration polish can fix. The incentive math is unforgiving: every party in the chain has arranged its exposure to be smaller than the exposure it is creating for the merchant. Here is the angle I think most coverage is missing. Everyone is reading the Visa TAP and Mastercard Agent Pay announcements as the beginning of a liability framework. I read them as the beginning of a fee layer. Think about the incentive structure the way I would think about a market maker's obligations. A payment network that can identify agent-originated traffic has created a new and valuable signal. It can sell that signal — as authentication, as protection, as agent purchase protection. What it has not done is accept the residual risk. Identification and indemnification are two very different products. The first is cheap to produce and easy to monetize. The second requires the network to hold capital against a class of errors it did not cause and cannot fully model. So the rational move for a network is exactly what we are seeing: build the identification layer, productize the protection layer as a paid add-on, and leave the tail risk where it currently sits, on the merchant. Amex's Agent Purchase Protection is a signal of intent, but until we see its coverage scope, its claim conditions, and its price, it is a marketing posture, not a risk transfer. I keep returning to the DeFi parallel because it is the cleanest precedent we have. When a protocol subsidized liquidity, the TVL chart looked like adoption. It was not adoption. It was a subsidy wearing a growth curve, and when the subsidy stopped, the liquidity left within hours. Agentic Checkout has the same structure if the liability is not priced. The eightfold traffic number looks like demand. If it is really merchants absorbing unpriced agent error, then it is a subsidy too — a subsidy paid by the merchant's returns line, and it will stop the moment the loss shows up in the P&L. The traffic is real. The economics are borrowed. And notice who is not at the table bearing risk. Constructor and Stripe expand the transaction endpoint and externalize the liability. That is not malfeasance. It is exactly what platforms do, and it is exactly what I would expect from a rational actor. But it means the welcome-agents posture is being sold to merchants by parties who have arranged their own exposure to be minimal. The counter-intuitive conclusion: the merchant is not choosing between welcoming and blocking agents. The merchant is choosing between two ways of losing money — losing it to foregone conversion, or losing it to agent error it cannot price. Silence is the loudest audit, and right now the silence on the liability question is deafening. No one has published the loss experience. No one has published the dispute rate. Until someone does, every merchant enabling Agentic Checkout is underwriting an unpriced risk with its own balance sheet. There is a second, quieter contrarian point. The Journal's three postures — QVC welcoming, Kate Spade browsing, Amazon blocking — are usually read as a maturity spectrum. I read them as a risk-tolerance spectrum. Amazon is not less sophisticated than QVC. Amazon is more exposed to agent error at scale, and it has the data to know it. The brands that block may be the ones who have actually run the numbers. Art burns hot; patience burns colder. The retailers waiting are not behind. They are pricing. I see the pattern before the price does, and the pattern here is a system closing at the technical layer while remaining open at the institutional and economic layers. The technology is ready. The rules are not. That is the whole thesis. So here is what I will be watching, and I would suggest any merchant or allocator watch the same. Whether the Cart Context specification ships and is actually usable for purchase-intent verification, because without it there is no way to prove an agent acted against the buyer's intent, and without that proof there is no liability assignment. Whether Visa TAP and Mastercard Agent Pay graduate from identification signals to genuine liability reassignment, or calcify into a fee layer. Whether a standardized agent-error reason code appears, because a reason code is the precondition for any dispute framework. Whether Amex's Agent Purchase Protection publishes coverage scope, claim conditions, and cost. Whether merchant structured-data coverage climbs above 15% and traffic-classification above 23%. And, most important, whether anyone publishes agent-driven return, chargeback, and dispute rates at all. The deeper question is who the customer actually is once an agent is buying. In every prior commerce model, the person at the checkout was the person bearing the consequence of the choice. Agentic checkout breaks that link. The agent chooses, the merchant fulfills, and the human discovers the outcome later. That is not a checkout upgrade. That is a separation of decision from consequence, and every market I have studied prices that separation eventually. The question is only whether the price is set by design or by disaster. Flows change, but the current remains. The current here is simple: capital does not flow into unpriced risk forever. If the liability question is not answered, merchants will not retreat all the way to blocking. They will retreat to the middle posture — browsing without buying. And that middle posture is where agentic commerce quietly dies: not with a ban, but with a locked door that everyone agreed to keep closed. The technology is ready. The institutions and the economics are not. That is not a product problem. That is a settlement problem. And settlement problems are solved with rules, not with APIs.