After agents fully replace apps

The entrance moves to whoever understands the intent

Portals, search engines and apps each changed how people reached a service. Portals organized content. Search organized routes to answers. Apps packaged a business into an interface people could learn and use repeatedly. Each shift redistributed attention, customer relationships and the ability to charge for access.

Apps still leave much of the coordination to people. To get something done, a user must know which services to find, in what order to use them and how to bridge the gaps. A flight, hotel and airport transfer may be three orders. For the traveler, they serve one purpose: arriving without trouble.

An agent takes over that coordination. It retains an understanding of the person, turns a goal into arrangements, chooses suppliers and keeps working when circumstances change. An agent that knows your travel preferences has no reason to forget your hotel or arrival time once the flight is booked.

Our view is that standalone apps will eventually cease to be necessary as the entrance to a service. People will still inspect maps, choose photographs, revise designs and enjoy browsing products. Those interactions can appear around the task at hand. Brands, inventory systems and fulfillment networks can endure without requiring every customer to enter a separate app.

Complete replacement has a demanding test. If users still have to work through a merchant’s prescribed process themselves, moving that process into a chat window has not completed the substitution. Games, creative work and social interaction are particularly difficult cases because participation is part of the value. The thesis has to hold there too; buying groceries cannot establish it. We expect agents eventually to organize those experiences as well, but that remains the hardest part of the argument to test.

Retailers will have agents too

Return to the camping trip. A personal agent can assemble a shopping list. Getting those goods to a collection point on your route requires the retailer to reorganize its own resources.

Which store has the stock? What can be substituted? Which delivery has space? When can the collection point accept the goods? What happens if you leave an hour later? The supplier must make decisions as conditions change and stand behind the resulting promise.

We therefore expect retailers such as Walmart to deliver services through their own agents. APIs will continue to carry inventory, prices and orders. Agents will decide how to use those capabilities to fulfill a customer’s objective. Conventional optimization software can handle scheduling and checks; the agents need not conduct every exchange in natural language. What matters is the retailer’s ability to assemble and revise a complete offer.

Walmart already describes an agent structure for customers, employees, partners and developers. Sparky uses multiple agents to coordinate customer assistance, while supply chain workflows use agents to position inventory and arrange delivery.^[1] Its January 2026 announcement with Google also used camping equipment as an example and set out plans to connect purchase history and delivery capabilities to Gemini.^[2] Collection at a point already on your route is the fuller scenario we propose here, rather than a service established by that announcement.

Walmart’s management has also said it does not expect all shopping to become automated, because people enjoy shopping.^[3] How much a customer wants to participate is a different question from whether a merchant needs a standalone app. An agent can leave the enjoyable choices to the person while taking over the work around them.

A fraction’s value depends on whom it reaches

A room left empty tonight loses tonight’s value tomorrow. A truck already on the road may still have a little space. An item that will not sell in its usual category may solve a particular person’s problem.

The internet has already made these resources easier to find. Airline yield management began exploiting differences in demand and the perishability of seats well before mass internet adoption. Research on online booksellers later showed how greater product variety could itself benefit consumers.^[4][5] Too Good To Go matches surplus food with buyers who accept a specified collection window. AWS Spot makes spare compute available to workloads that can tolerate interruption.^[6][7] Each finds buyers willing to accept particular conditions.

Agents can make those matches more precise. Your agent knows you can leave a day later but cannot make another connection. You can change brands but cannot accept food containing nuts. You will collect goods on your route but will not drive half an hour to save a small amount. Those tradeoffs create new uses and combinations for the same inventory.

The camping supplies illustrate the point. A bag of ice, a few meals and some delivery capacity are ordinary resources in isolation. Put them in the right place, at the right time, along tomorrow’s route, and preparing for the trip becomes easy.

The additional value comes from putting resources to a better use. If a merchant merely becomes better at estimating what a customer will pay and charges more for the same order, the main change is how the surplus is divided. Personal knowledge can improve service or raise prices. These outcomes should be counted separately.

Why an empty airline seat might be given away

Suppose a flight has covered its costs but still has empty seats shortly before departure. Putting the right people in those seats may be worth more than leaving them empty.

Travel agents have long used personal knowledge of clients to match people with available products. An airline’s AI agent could negotiate with travelers’ agents: who wants to visit this destination, who can accept this departure time, and who might become a lasting customer after a good experience? The offer could be a cheaper ticket, an upgrade or even a gift.

Covering a flight’s cost does not, by itself, make a free ticket worthwhile. Another passenger may bring taxes, service and baggage costs. The seat might still sell at a normal fare. Teaching customers to wait for gifts could also damage future revenue. The relevant comparison is whether additional revenue and customer value exceed the added costs, the sale forgone and the effect on normal bookings. Conversely, a flight that has not covered its total cost may still benefit from a low-fare booking that makes a positive contribution.^[4]

Agents make the conditions easier to identify and act on. A traveler willing to accept late confirmation, limited dates or particular baggage terms may qualify for an offer that would make little sense as a universal fare cut. Giving seats away can be useful in some circumstances; it is not the best answer for every empty seat.

The fraction here is a service opportunity about to expire. Knowing who needs it and can accept its constraints creates a chance to turn waste into an experience. Unsold rooms, appointment gaps and idle equipment hours follow a similar logic.

More small tasks become worth doing

Many transactions never happen because arranging them takes too much effort. Saving a little money takes half an hour of comparison. Recovering a small fee requires several forms. Coordinating a trip for ten people takes repeated rounds of confirmation. The potential benefit exists, but may not justify the person’s time.

In his 1999 discussion of micropayments, Nick Szabo pointed out that deciding whether a transaction is worthwhile itself consumes attention.^[8] Agents extend this problem into execution. Once they find an opportunity, they must still compare, pay, track, amend and sometimes pursue a refund.

When the coordination cost saved exceeds the cost of computing, checking and correcting mistakes, previously abandoned tasks become feasible. The boundary of the market moves. A transaction can be small, provided it remains worth delegating.

Anthropic’s Project Deal offers a limited but concrete example. In December 2025, 69 employees let agents represent them in buying and selling personal belongings. Of four parallel runs, the one chosen for real exchange produced 186 deals worth just over $4,000. Each participant received a $100 experimental budget, so the result does not establish that such a marketplace can sustain itself commercially. It does show agents turning different human preferences into real trades.^[9]

The next test comes when subsidies disappear, participants are strangers, and refunds and disputes arrive. Whether the service remains worth using will determine how far it spreads. The important cost is that of completing a task successfully, including the work after the first answer.

Whose interests does the agent serve

As service becomes more personal, the agent’s incentives matter more.

Sellers want higher revenue. Buyers want suitable goods on suitable terms. An agent paid by a merchant while presenting itself as the customer’s independent adviser may blur those goals. The FTC’s preliminary surveillance-pricing study describes how personal data can influence prices and product presentation. Its examples were hypothetical to protect commercial confidentiality, rather than findings of wrongdoing by named firms.^[10]

A buyer’s agent could also shift the balance. It can keep comparing alternatives, remember renewal conditions, apply for discounts and decline unnecessary upgrades. To earn that role, it should explain its choices, disclose commissions and let people take their preferences and transaction records elsewhere.

The business model faces a constraint too. Eric Seufert’s criticism of agentic commerce on X centers on thin retailer margins and the burden of another commission.^[11] If an agent simply reroutes an existing order and adds a fee, neither side may want to keep paying. Creating transactions that otherwise would not happen, or saving real fulfillment and service costs, gives the intermediary a more durable reason to exist.

Google’s Universal Commerce Protocol is already attempting to establish a common language for discovery, purchasing and support between agents and merchants.^[12] Protocols can help connect the parties. Loyalty, fees and responsibility still have to be worked out. As agents move closer to the center of commerce, those questions become harder to avoid.

From clicks to continuing instructions

In the app era, a business interaction often begins when someone opens an interface. With agents, a person can leave a standing instruction: arrange a trip if a suitable flight appears next month; keep a bill below a certain level; have the camping supplies ready before departure.

Those objectives can trigger continuing searches, comparisons, negotiations and actions in the background. One human intention can produce many exchanges between machines. This is the implication for computing infrastructure: some of its work becomes directly connected to fulfillment, savings and the use of physical resources.

Demand will not grow without limit merely because calls multiply. A direct lookup need not become repeated inference. Recurring tasks can reuse results. Error recovery can consume the profit. Computing capacity is worth building when it can complete valuable delegated work at a viable cost.

To judge this transition, we would ask: how many tasks still force people back into apps? How much inventory gets used because of new matches? After all agent and fulfillment costs, do both sides want to trade again?

Our expectation is that agents will replace apps completely. People will begin with their own intentions, and suppliers will deal with agents able to understand and carry them out. Fractions that have always existed, but never met the right demand, will be found again.

When you set off tomorrow, the supplies are already waiting on your route. The seat that would have flown empty has found someone willing to take the trip.