The Interface Was the Moat

Something happened to experienced programmers over the last few years that almost nobody outside the field noticed, and it is about to happen to everyone who buys anything.

If you are a genuinely good programmer with ten or twenty years behind you, the modern models have handed you something specific. You can now write in any language. Not eventually, not after a weekend of tutorials. Today. The syntax is no longer a gate. The framework conventions are no longer a gate. You sit down and you write, and the nuances of that particular language get discovered through iteration rather than through a month of reading.

What that revealed is uncomfortable and clarifying at the same time. The syntax was never the skill. Knowing where the semicolons go was a tax you paid to get access to the thing that actually mattered, which was knowing how to decompose a problem, where state should live, what will break at scale, and which abstraction will still make sense in two years. Those did not get automated. The tax did.

Consumers have been paying an almost identical tax for thirty years. It is about to be refunded, and the refund is going to cost somebody a great deal of money.

The tax consumers pay

Every retailer built its own interface. Its own cart. Its own account system, its own checkout flow, its own return policy, its own sizing guide that agrees with nobody else's sizing guide. To buy something you first had to learn the shop.

That is a real cost and people felt it, even if they never named it.

Amazon's actual product was not selection. It was that you learned one interface, one checkout, one return policy, and then never had to learn another. Hundreds of retailers collapsed into a single set of habits. People moved there in enormous numbers and the thing they were voting for was not price. It was not having to think.

Food ran the same play a decade later. Consider ordering a pizza. The independent place a mile away has a website and a cart and a menu you have to decode. The chains each have their own app with their own login and their own deal structure. A place slightly further out has something worse. Each one is a small separate act of learning, and the friction is enough that people default to whatever they already know.

Delivery apps consolidated exactly that. One app, one saved card, every menu in the area, one set of rules. DoorDash got there first and Uber Eats followed, which made obvious sense because the drivers were already moving. And consumers did something worth paying attention to: they accepted a meaningful price premium for it. Fees, inflated menu prices, a service charge. They paid because the alternative was learning twelve interfaces.

The pattern across both waves is the same. People gravitate to whatever removes the need to learn, and they will pay for it.

What agentic commerce removes

Now put the programmer transition and the consumer transition side by side, because they are the same transition.

The consumer already knows how to shop. They know what they want. What they have been forced to carry is the interface layer: the account, the cart, the checkout, the return flow, the loyalty scheme, the app that needs updating.

With a capable agent and a payment method it can actually use, that layer disappears. You say you want a pepperoni pizza from Goodfellas in Marietta, or from one of the chains, or from the place across town you have never ordered from before, and it happens. You never see a website. You never learn a cart. You never read a return policy.

The symmetry with programming is exact. The programmer no longer learns the language. The consumer no longer learns the store. In both cases the thing that got removed was never the valuable part. It was the toll.

The mechanism that makes this real rather than a demo is agentic payments: an agent that can be handed spending authority in a scoped, revocable, auditable way. That is the piece the whole shift waits on, and it is being built now.

The stored card was the other half of the lock-in

The interface is only half of what the aggregators took. The other half was the card on file.

It sounds trivial and it is not. Amazon built one-click checkout because re-entering a card is enough friction to lose a sale. Every delivery app's real retention mechanism was that your card was already in it. That is why people defaulted to the app they had already set up, even when the restaurant's own site was cheaper.

So what happens to the card when an agent does the buying?

Not much, if the agent is just typing your card number into checkout forms on your behalf. That is the naive version and it inherits every problem card rails have with machines: chargebacks written for humans, fraud scoring that treats automation as suspicious, and no clean way to scope or revoke authority for a single purchase.

The version being actually built looks different. The agent carries its own funded balance and settles directly, and both of the serious attempts at this revive the same long-dormant HTTP status code: 402, Payment Required, which has sat unused in the spec since the early web.

The shape is the same in both. An agent requests a resource. The server answers 402 with payment instructions. The agent pays and retries with proof attached. No account, no login, no stored card, settlement in seconds.

There are two competing standards and it is worth knowing both, because a merchant does not get to pick which one their customer's agent speaks.

x402, from Coinbase. Stablecoin-native, most traffic settling in USDC on Base or Solana. Coinbase and Cloudflare have moved it under a foundation with Circle, Stripe and AWS involved, Stripe shipped support in February 2026, and it is one of the first extensions to Google's Agent Payments Protocol. It currently carries the most volume of the two.

MPP, the Machine Payments Protocol, from Stripe and Tempo. Specification at mpp.dev, launched March 2026 alongside Tempo's mainnet. Two things make it notable. It is not stablecoin-only: Stripe offramps agent stablecoin payments into an ordinary Stripe balance, and fiat and card payments run through it via Shared Payment Tokens, with Visa having published card specifications and an SDK for exactly that. And it introduces a sessions primitive, where an agent authorises a spending limit once and then streams many small payments without a separate settlement per interaction.

So the honest framing is not "stablecoins replace cards." It is that machine payments need three properties card rails were never designed for: small amounts, high frequency, and authority that is scoped and revocable per purchase rather than per account. Stablecoin rails got there first because they could. The card networks are arriving through the same protocols rather than being displaced by them.

What that actually requires of a merchant

This is the part that turns a macro trend into a to-do list, and it is longer than most operators expect. To sell to an agent you need:

1. A machine-readable menu or catalogue. Items, prices, options, real availability, real constraints. Not a PDF. Not a pretty page.
2. An endpoint that can quote and accept payment in whatever the agent is carrying, which increasingly means stablecoin settlement rather than a card form.
3. Order notification into your existing flow, so a paid agent order arrives where your staff already look rather than in a system nobody watches during a rush.
4. A channel back to the agent. This is the one everybody forgets. If an item is out, if the order is wrong, if the customer wants a refund, there is no human on the other end to phone. There is an agent, and it needs a defined way to be told.
5. Discoverability in whatever directory agents actually consult. Early registries exist: x402scan and MPPscan, indexes for the two standards, both maintained by Merit Systems, who also build AgentCash, a payment and discovery layer that lets an agent pay per call from a single balance.

Look at that list again and notice what it is. It is a full integration project against two competing standards, and almost no independent restaurant or small retailer can do it alone.

That last clause is the whole thing. A pizzeria owner cannot evaluate settlement protocols. They could not build a delivery app either, which is precisely why they ended up renting one at 30%. The same gap is opening again, and the question is whether it gets filled by another rent-seeking aggregator or by something the merchant actually owns.

The honest state of it

This is early and the narrative is well ahead of the volume. In March 2026, onchain data showed x402 processing on the order of tens of thousands of dollars a day, a great deal of it testing, against an ecosystem valued in the billions. Anyone telling you agentic commerce is here today is selling something.

But the protocol works, the facilitators are live, and the companies who own the payment rails are building it rather than dismissing it. The gap between "works" and "widely used" is a question of years, not of whether.

The part the aggregators are not going to enjoy

Here is the unintended consequence, and it is worth stating plainly.

A delivery app's moat was the interface, not the logistics.

The valuable thing DoorDash built was that consumers learned it and stopped learning anything else. The menus, the saved card, the habit. The fleet of drivers was real and expensive, but it was never the defensible part, because a fleet is assembled with an app and driver economics that are good enough. That is a hard operational job. It is not a moat. Several companies have built one independently, which is the definition of not being a moat.

If the consumer stops opening the app, what remains is the commodity half of the business.

The demand-side aggregation was the asset and the agent dissolves it. The logistics survives, because somebody still has to physically carry the food, and that need is not going anywhere. But "we move things" is a margin business that competes on price and coverage. "Everyone in the country has our icon on their home screen and their card on file" was the actual franchise.

This does not happen next quarter. It happens the way these things always happen, which is slowly and then in a rush, once one agent does it well enough that trying it feels obvious.

The arithmetic, and who it actually helps

This is the part I find most persuasive, and it is the part getting the least attention.

Marketplace commissions on the delivery platforms are tiered. As of 2026 they run roughly 15% to 30% of order value depending on plan, and pickup orders that use no driver at all still take around 6% or 7%. Uber Eats raised its rates in March 2026. Once payment processing, sponsored placement and funded promotions stack on top, independent operators commonly report a blended effective cost somewhere between 25% and 35% of gross order value. Rates vary by city, tier and contract, so treat those as planning baselines rather than quotes.

Now the contrast that makes the whole argument. The same companies already sell commission-free direct ordering. DoorDash Storefront and Uber's Webshop both charge essentially payment processing and nothing else, in the region of 2.9% plus 30 cents and 2.5% plus 29 cents respectively.

So the difference between an order that arrives through the marketplace and the identical order arriving through the merchant's own page is already something like fifteen to thirty points of revenue. In an industry where net margins are frequently single digits, that is not a line item. That is the business.

And you are already paying for it. Merchants raise menu prices on the apps to absorb the commission, which is standard practice and openly permitted. So the current arrangement is that the consumer pays an inflated menu price, plus a delivery fee, plus a service fee, and a double-digit share of that stack funds the layer whose contribution was the interface.

Remove the need for the interface and that rake has no reason to exist.

If the agent goes to the merchant's own ordering page and pays through the merchant's own processor, the transaction costs roughly 3% instead of roughly 30%. Delivery still costs money, because somebody still has to drive. Ordering becomes close to free.

That is a genuinely deflationary force in a sector that has been raising prices for years while operating on thin margins, and it has three winners at once:

• The merchant, who stops paying rent on customers they already had.
• The consumer, who stops paying the inflated price that funded that rent.
• The wider economy, because this is a large, low-margin, high-employment sector and the saving is structural rather than a one-off discount.

The mom-and-pop operator is the biggest beneficiary here, which is not how these transitions usually go. Independents had the least leverage to negotiate commissions and the least ability to absorb them. They are the ones who get the most back.

The consequence nobody is pricing in

Each of these consolidation waves was sold as convenience, and each one ended with a toll booth in a new place.

Retail fragmentation ended with Amazon collecting a fee and owning the customer. Restaurant fragmentation ended with the delivery apps collecting a fee and owning the customer. Both were genuinely better for consumers. Both also transferred the customer relationship away from the merchant, permanently, and then raised the rent once the relationship was secure.

So the obvious worry is that agentic commerce is just the next iteration, and the rake reappears somewhere else.

I think that is half right, and the half it gets wrong is the important half. It depends entirely on one thing: whether the customer names the merchant.

Named intent. "Order a pepperoni pizza from Goodfellas." There is nothing for an intermediary to decide. The agent goes to that merchant, orders from that merchant, pays that merchant's processor. Nobody is owed a finder's fee for finding something that was already named. This case disintermediates completely, and it is the case in the arithmetic above.

Unnamed intent. "Order me a pepperoni pizza." Now something has to choose, and that choice is worth money. Whoever makes it will eventually charge for making it, whether as placement, ranking or a cut. That is the new toll booth, it sits at the moment of intent rather than the moment of transaction, and it is being built right now mostly by companies that are not restaurants.

So the forecast is not that the rake disappears. It is that the rake collapses for named orders and relocates to discovery for unnamed ones. Which produces a conclusion I did not expect when I started thinking about this: brand gets more valuable, not less. Being the restaurant a customer asks for by name is the difference between paying 3% and paying whatever the discovery layer decides to charge.

The other quiet consequence: if the agent buys, nobody sees your brand. No homepage, no photography, no carefully built checkout, no upsell, no loyalty prompt. Your restaurant becomes a structured feed of items, prices, availability and constraints. Everything you spent money on to differentiate the experience is invisible to the thing doing the buying.

What a merchant should actually do about it

Not panic, and not wait. Four things, in order of how cheap they are.

Be machine-readable. Accurate, structured, current menu and inventory data, with real availability and real constraints. An agent cannot order what it cannot parse. Merchants whose data is a mess will simply not appear, and they will never receive a single signal explaining why.

Make sure you have a direct ordering path that actually works. This is the one that pays for itself immediately, before any of this arrives. If the agent has somewhere to go that is not the marketplace, the fifteen to thirty points stay with you. If it does not, the agent will use the marketplace and you will keep paying.

Invest in being asked for by name. This is the conclusion that surprised me. Named intent is free and unnamed intent will be sold. Every dollar that makes a customer say your name instead of saying "a pizza" is now worth considerably more than it was, and that is the opposite of the usual advice that brand stops mattering when a machine is buying.

Decide what to do with the saving. A merchant who was raising prices to cover a 30% commission can stop raising them, or can keep the margin. That is a real strategic choice and it is worth making deliberately rather than by default.

The programmers who did well out of their transition were the ones who already had the fundamentals, and simply let the tax fall away. This looks like the same shape of event, and unusually it favours the small operator: the merchant who never had the leverage to negotiate a commission gets the biggest proportional refund.

The layer that is exposed is the one whose product was the interface. It turns out that was the thing being sold all along, and it was never the driving.