Constructor and Stripe Closed the Checkout Gap. The Liability Ledger Is Still Empty.

Evidence suggests the technical layer of agentic commerce is now closed. The institutional layer is not. On October 5, Constructor and Stripe pushed an agentic checkout endpoint into production. A configured merchant can now let a shopping agent — not a human — complete a purchase inside a conversation thread. Link by Stripe settles the payment. The merchant does not rebuild its payment stack. It flips a configuration flag. That is the entire event, and it is not a product review. It is a structural change. The transaction endpoint has moved from the browser interface to the AI agent itself. Discovery and purchase no longer sit on either side of a checkout page. They collapse into a single call. The structural consequence is that the merchant's last human checkpoint disappears. Historically, a checkout page was a control surface a human could cancel. An agentic endpoint removes that checkpoint from the merchant's view. The agent holds the state. The merchant receives an order. I have spent the past weeks reading the same integration documentation every enterprise architect will read this quarter. The code path is deterministic. The settlement path is not. That asymmetry is the subject of this analysis. The architecture splits into four layers, and each layer reports to a different owner. The application layer belongs to Constructor. The Commerce Reasoning Engine generates intent. The AI Shopping Agent executes discovery and recommendation. The AI Product Insights Agent handles natural-language query. These components decide what a buyer might want. The transaction layer belongs to Stripe. Link by Stripe executes payment inside the conversation. Merchants already on Stripe require only a configuration change. Merchants outside Stripe route through Constructor's own Stripe integration. They do not migrate their primary processor. Constructor removed the switching cost that usually kills adoption. The network and standards layer is fragmented. Visa TAP and Mastercard Agent Pay emit signals that identify agent-initiated traffic. American Express shipped Agent Purchase Protection in April 2026. The Cart Context specification — the document that would define how purchase intent is verified — remains in draft. The merchant and platform layer is openly divided. Sephora, Under Armour, and The Very Group are live. The Wall Street Journal described three postures: QVC welcomes agents, Kate Spade permits browsing but blocks purchase, Amazon blocks entirely. Shopify defaults to agent-ready. The three postures are not preferences. They are risk prices. QVC welcomes agents because its catalog is shallow and its return rate is tolerable. Kate Spade blocks purchase because apparel has a size and fit failure mode an agent cannot verify. Amazon blocks because it operates a competing agent and will not route demand through a third party's reasoning engine. The first gap is readiness, and the numbers are not ambiguous. Only 11% of small and mid-sized businesses meet an agent-ready standard. Only 15% of merchants hold the structured product data an agent needs to select a SKU. Only 23% can distinguish AI traffic from human traffic. Shopify's default-on policy drove an eightfold increase in AI-driven traffic in Q1 2026. Read those four figures together. The enabling cost collapsed while operational capacity did not. A merchant can accept an agent's purchase with one configuration change. That same merchant cannot verify intent, cannot classify the session, and in most cases cannot feed the agent a machine-readable catalog. Integration is trivial. Absorption is not. The eightfold traffic figure is the tell. Shopify's default did not create demand; it removed friction. The same mechanism will surface agent error at scale, and the 23% classification rate means most merchants will not see it until chargebacks arrive. I have seen this asymmetry before. In 2020, while finalizing my thesis on formal verification, I audited the early Curve stablecoin pools and found three integer overflow vulnerabilities in the pre-launch documentation. The math libraries were elegant. The implementation checks were absent. Here, Constructor has shipped a clean transaction path into a merchant base that lacks the instrumentation to monitor it. Trust is a variable; proof is a constant. The second gap is liability, and it is the actual finding. 93% of merchants believe the AI provider or agent operator should absorb the financial loss when an agent executes a wrong purchase. The current model assigns 100% of that risk to the merchant. An agent selects the wrong size, the wrong color, the wrong SKU. The cost lands on the merchant's balance sheet. The payment infrastructure cannot yet separate a legitimate human-initiated purchase from an agent-initiated transaction that fails to match buyer intent. There is no standardized reason code for agent error. Visa TAP and Mastercard Agent Pay identify agent traffic. They do not reallocate responsibility for non-fraudulent mistakes. Amex's Agent Purchase Protection and Cart Context are early signals. Neither has a ratified specification. A reason code is not a formality. It is the accounting primitive that lets a dispute be priced, routed, and settled. The absence of an agent-error code means agent disputes fall into existing human-error buckets, which default to the merchant. The conclusion is mechanical. The networks can see that an agent acted. They have not agreed to price what happens when the agent acts wrongly. Default risk stays with the merchant. The third gap is unit economics, and it remains unverified. A merchant that enables Agentic Checkout is underwriting a bet: that Constructor's reasoning engine produces purchases the customer actually wanted. If the bet holds, the welcome posture looks prescient. If it fails, the merchant absorbs returns, disputes, and chargebacks with no counterparty to bill. The technical path is ready. The per-transaction economics are not. I trace this the way I traced the Anchor Protocol yield contracts in 2022. The headline metric was inflows. The reality was that the yield was unbacked debt, not revenue. Here the headline metric is conversion. The unexamined metric is the cost of agent error, and nobody has booked it. The bulls have one argument that survives scrutiny. Low enablement cost is the correct strategy precisely because it forces the liability question into the open. If Constructor had required a payment-stack migration, adoption would have stalled at a handful of enterprise pilots. By removing the switching cost, Constructor pushed thousands of merchants into live exposure. That is uncomfortable, and it is useful. The 93% survey figure becomes an operational demand rather than a survey opinion. The identification layer is also the correct first step, even though it is incomplete. Visa TAP and Mastercard Agent Pay cannot reallocate liability until they can classify the counterparty. A signal precedes a rule. The sequence is rational. What the bulls miss is the direction of the incentive. Payment networks monetize identification. Protection products are a revenue layer. There is no obvious commercial reason for a network to volunteer liability for non-fraudulent agent error. The gap does not close on its own. It closes only when merchants refuse the default or when a regulator forces the allocation. I traced 14 wallet clusters linked to SBF's personal accounts after the FTX collapse. The lesson was not that opacity existed. It was that opacity persisted because every intermediary found the default profitable. The forward question is narrow. When the Cart Context specification ratifies, will it define intent verification, or only agent identification? If it defines only identification, the merchant remains the residual risk-bearer and the welcome posture is financially fragile. Watch six metrics: ratification of Cart Context, whether TAP and Agent Pay extend to liability reallocation, Amex claim terms, the appearance of a standardized agent-error reason code, merchant data readiness above 15%, and the return and chargeback rate on agent-driven transactions. Trust is a variable; proof is a constant. The rail is built. The proof is not.