On Agentic Commerce, Part I: Intent

I think most of the current conversation around agentic commerce starts too late in the transaction.

The category is usually described through payments, checkout, product feeds, browser automation, or machine-to-machine purchasing: can an AI discover a product, authenticate a payment method, call a paid service, and complete a purchase without a human clicking the final button? Those are meaningful changes, and a great deal of infrastructure is being built around them. But in most of those cases the agent is operating more efficiently inside a market whose structure already exists. The product has already been defined, the seller has decided what it costs, inventory and fulfillment have been encoded into software, and the buyer is choosing among known options. AI is improving how the user reaches the transaction; it is not necessarily changing how the transaction itself is formed.

I think the more consequential shift is further upstream. For most economic activity, the fundamental input is not a SKU, merchant, payment instruction, or even a search query. It is some state of the world that a person or business wants to change. Someone needs a mattress that will not aggravate a bad shoulder. A company needs twenty qualified creators for a launch next month. A founder needs an attorney with specific regulatory experience this week. A procurement team needs a component produced at a particular quality, volume, cost, and deadline. The transaction is downstream of that intent. Historically, people have been responsible for translating the intent into the structure required by markets and software. Agentic commerce begins to matter when software can perform that translation itself.

This is why I think the defining interface of agentic commerce will be intent.

What I mean by agentic commerce

At the broadest level, agentic commerce is economic activity in which AI materially participates in deciding, forming, or executing an exchange. That definition intentionally includes much more than autonomous payments. There is a progression from an agent helping a human make a decision, to initiating an action after approval, to executing within delegated authority, to a world in which agents represent both sides of an economic relationship and continuously decide whether, when, how, and with whom to transact.

We are still early in that progression. Much of consumer agentic commerce today is better described as agent-assisted or agent-initiated: a user expresses what they want, an AI helps clarify the request and research the available options, and the existing commerce stack does most of the rest. ChatGPT shopping is a useful example. The model can reason over a richer description of the buyer’s preferences than a traditional search engine can, while merchant feeds, structured catalogs, ranking systems, checkout infrastructure, and payment rails still provide the market underneath it. Shopify can make enormous amounts of merchant supply legible to an AI surface without every merchant independently rebuilding its commerce stack for that interface.

That is already a meaningful improvement, but it is the relatively easy case because the supply has already been structured. The harder part begins when the system must form the transaction around the intent rather than select a pre-existing transaction from a catalog. That distinction separates two futures that are frequently bundled together under the same term: one in which AI makes access to the existing economy dramatically easier, and another in which AI begins reorganizing economic activity around what people and businesses actually want. I expect the former to become ubiquitous. The latter is where I think the structure of markets begins to change.

Search is an artifact of software that could not understand us

For most of computing history, humans have adapted themselves to the structure of machines. Databases required fields; software required menus; marketplaces required categories; search engines required queries; travel sites required dates, locations, cabin classes, and passenger counts; enterprise systems required workflows defined in advance. The user learned the representation the system expected and translated whatever they actually wanted into it.

Search was an extraordinary solution to this constraint. Rather than requiring people to know where every piece of information lived, search engines indexed enormous amounts of heterogeneous supply and gave us a lightweight language for navigating it. Marketplaces did something similar for commerce: they normalized sellers into listings, introduced filters and reputation, aggregated demand, and dramatically reduced the cost of finding the other side of a transaction. A great deal of the internet economy was created by making previously messy supply searchable.

But search is still translation work performed by the user. If I say that I sleep on my side, wake up with shoulder pain, dislike mattresses that retain heat, want to be able to try one at home, and have a $1,200 budget, I have already described the actual decision problem. Historically I would turn that problem into several queries, read reviews, inspect return policies, compare materials, decide which sources I trusted, and reconstruct the answer myself. The relevant information may already exist on the internet; the friction is converting my actual preference structure into something the market can resolve.

Large language models reverse that interface. They can accept the messy representation first and derive the structured work afterward. A user no longer has to begin by knowing the category, merchant, query, workflow, or sometimes even the exact product they need. They can begin with the desired state. This sounds superficially like a change from a search box to a chat box, but the deeper shift is that if the interface starts with intent, the software behind it has to become capable of resolving intent. That is a substantially larger job than returning results.

Intent is not a prompt

The word intent can become vague very quickly, so it is worth being precise. A query asks for information. An instruction specifies an action. A goal specifies a desired state. Intent is the broader collection of objectives, constraints, preferences, tradeoffs, relationships, context, and authority that determines whether an outcome is actually good.

“Flights to Tokyo” is a query. “Book United 837 tomorrow” is an instruction. “Get me to Tokyo before noon Tuesday for less than $2,000” is a goal. The intent behind it may include things I never state in the current prompt: I strongly prefer direct flights, value arriving rested more than saving $150, avoid certain airports, care about reliable Wi-Fi, and am willing to spend more when the trip is unusually important. A capable system should not require me to reconstruct that state from scratch every time I ask it to do something.

This becomes more important as agents become persistent. People and businesses do not experience their economic lives as independent prompts. They have unresolved goals, recurring obligations, existing relationships, budgets, preferences, permissions, contracts, and context that persists over time. A useful agent should know which of those things matter, what it has already been authorized to do, what happened previously, and what remains unfinished.

A great deal of economically meaningful intent is also standing rather than episodic. A company wants inventory maintained above a threshold, qualified candidates continually surfaced for hard-to-fill roles, customer acquisition kept within a payback target, or cloud costs reduced without degrading performance. A buyer may want to know when a certain item becomes available below a price; a procurement team may want alternatives if a supplier becomes unreliable. None of these should require a human to remember to initiate the same search every morning. Once intent persists, discovery can become proactive: demand can remain in the system while supply changes, and an opportunity can appear because the states of the two sides suddenly become compatible.

That is a different economic primitive from search.

Listings describe supply. Goals describe the world a buyer wants.

The listing is arguably the foundational abstraction of internet commerce because it makes supply legible before a buyer arrives. A merchant specifies what exists, what it costs, how much is available, and under what conditions it can be purchased. Once enough supply has been normalized this way, search, ranking, comparison, payment, and fulfillment become tractable software problems.

This abstraction works exceptionally well for structured supply. A shoe can be represented by model, size, color, inventory, price, shipping, and return policy. A hotel can expose dates, room types, rates, availability, and amenities. Software can expose capabilities, pricing, usage limits, and endpoints. The economically relevant state of the thing exists independently enough from the buyer that it can be published in advance.

A large amount of economic supply does not work this way. Whether a lawyer is right for a matter depends on expertise, conflicts, availability, geography, urgency, budget, relationships, and the specifics of the request. A creator may accept one campaign and reject another at the same price because the brand, timing, audience, exclusivity, deliverables, or creative requirements differ. A manufacturer’s quote may depend on volume, tooling, material, tolerances, quality standards, payment terms, delivery schedule, and current utilization. A consultant can simultaneously be “available” in the abstract and unavailable for the only engagement a particular buyer cares about.

The difference is that the relevant supply state does not fully exist independently of demand. Part of the transaction has to be computed after the request arrives. That is why I think goals become increasingly important alongside listings. A listing says, “This is what I can offer.” A goal says, “This is what I need the world to make true.” For simple commerce, resolving the goal may mean retrieving one product from a catalog. For complex commerce, it may require discovering several possible counterparties, gathering dynamic or private state, negotiating different terms, creating conditional agreements, coordinating execution, and verifying the result.

From the user’s perspective the goal remains one object even when the economic machinery underneath it contains dozens of transactions. This is a subtle but important inversion: instead of forcing the buyer to decompose an outcome into markets and transactions before software can help, the agent begins with the outcome and becomes responsible for constructing the economic work required underneath it.

Checkout is not the transaction

A large amount of agentic-commerce infrastructure naturally focuses on payment because payment is one of the hardest pieces of existing software for autonomous systems to use safely. Machine-payment protocols make it possible for agents to programmatically purchase APIs, data, compute, browser sessions, and other software resources; commerce protocols are making product discovery, checkout, payment, fulfillment, and order state increasingly accessible to AI surfaces. These are necessary pieces of the stack.

But from the perspective of the underlying economic problem, payment is one state transition in a much larger lifecycle. Before money moves, someone may have to determine who is capable of satisfying the goal, establish whether they are available, evaluate trust, compare alternatives, disclose enough information to create a quote, negotiate price and non-price terms, coordinate multiple parties, and obtain authorization. After the payment, there may be fulfillment, dependencies, revisions, verification, refunds, disputes, support, recurring obligations, or a continuing relationship between the parties.

For a $20 API call, almost none of this matters. For a six-month services engagement, almost all of it matters. The economic value available to agents therefore expands as they can take responsibility for progressively more of the transaction lifecycle. An agent that can autonomously purchase a known product is useful. An agent that can take a loosely specified objective, determine what economic actors and capabilities are required, establish agreements with them, manage the execution, and return a verified result is something much closer to a new economic actor.

This leads to a broader view of agentic commerce than “agents buying things.” The long-term problem is economic coordination. Payments are part of that coordination, just as search is part of discovery, but neither is the full transaction.

The economy resolves through three execution domains

When thinking about the supply an agent ultimately acts against, I find it more useful to divide the economy by what actually has to execute the work: humans, digital systems, and the physical world. Terms like dynamic, multi-party, negotiated, recurring, and long-running describe the mechanics of a transaction, but each of those mechanics can appear across several domains. A human transaction can be highly structured; a software transaction can be negotiated; a physical transaction can involve dozens of parties and persist for years.

Human execution is the least structured and, for that reason, one of the hardest domains to make agentic. People carry dynamic capability, availability, willingness, price, preferences, permissions, relationships, conflicts, and reputation. A lawyer may be qualified but conflicted out. A creator may be available but unwilling to work with the brand. An evaluator may only be useful if they have a narrow combination of education, domain expertise, geography, language, or prior exposure. The economically relevant state cannot be fully represented by a static listing because whether the supply exists depends partly on the particular request.

This becomes even harder when one goal requires many people. “Get twenty qualified evaluators to review this system by Friday” should appear to the buyer as one objective, not twenty independent marketplace purchases. Underneath, however, the system may need to define qualification criteria, identify a much larger candidate pool, determine availability, communicate the work, reconcile different prices and terms, coordinate timing and instructions, verify completion, handle exceptions, and combine many outputs into one result. Creator marketing, AI evaluation, recruiting, research, and professional services appear to be different verticals, but many share this same structure: heterogeneous human supply must be formed around a buyer’s specific goal.

Digital execution is much more naturally machine-readable. APIs, software, models, inference, data, browser sessions, compute, storage, and other software capabilities can expose standardized interfaces, deterministic inputs and outputs, and programmatic pricing. Agents can compare several providers, invoke one, observe the result, and switch again in seconds. I expect this part of the economy to be especially open and competitive because interoperability and low switching costs strongly favor standard protocols. The differentiated question increasingly becomes not whether an agent can call a capability, but whether it knows which capability to call, how to compose several capabilities into a larger plan, and whether their outputs actually moved the underlying goal forward.

Physical execution covers outcomes that ultimately have to appear in the physical world: retail products, food, transportation, logistics, travel, manufacturing, property, equipment, and eventually much more robotic activity. Decades of ecommerce have already structured enormous portions of this supply into catalogs, inventory systems, booking systems, logistics software, and payment infrastructure, which is why physical commerce is one of the most visible early applications of agentic interfaces. But physical execution does not imply simple commerce. Custom manufacturing, logistics, construction, or complex travel can involve negotiated specifications, scarce capacity, conditional pricing, changing state, and many counterparties. The final output may be physical even when the transaction itself looks much more like a negotiated services market.

These three domains will increasingly collapse together from the user’s perspective. “Launch this product in Japan” may require human legal and creative work, digital translation and analytics systems, and physical manufacturing and logistics. The user should not have to decide which economic domain to enter or manually stitch together the providers. The user has one goal. The agent constructs the execution graph beneath it.

Communication belongs inside the economic graph

Not every economically meaningful interaction ends in a transaction, and a useful economic network should not pretend otherwise. Buyers ask questions, sellers clarify requirements, companies explore partnerships, customers request support, suppliers communicate changes in capacity, and relationships may develop for months before a purchase occurs. Communication is not incidental to a market; it is part of how markets form.

Open protocols will make this communication increasingly easy. Agents can expose identities and capabilities publicly, communicate through standardized interfaces, and remain technically reachable without belonging to the same proprietary system. That is a good thing. But the ability to send a message is different from possessing the context that makes the message economically useful. Knowing that an endpoint exists does not answer which agent I should contact, whether it is authorized to make an agreement, what happened the last five times our principals worked together, what terms they have historically accepted, how reliable they have been, or whether their representation of the underlying person or business has changed materially.

Those are questions about shared state. The communication itself can be open and inexpensive while the context around it remains valuable. I therefore expect basic agent-to-agent messaging to commoditize much faster than economic identity, relationships, reputation, authority, and outcome history. The protocol tells two agents how to talk. The network can tell them whether they should.

The open agent web does not eliminate networks

There is an intuitive argument that sufficiently capable agents make closed economic networks less important. If every business exposes an A2A endpoint, every merchant supports an open commerce protocol, every machine can accept programmatic payment, and general-purpose models can crawl and reason over the entire web, an agent could theoretically assemble its own economic network dynamically from public endpoints. Why should another network exist in the middle?

I think this confuses connectivity with network value. The open web made almost any document addressable; it did not make every document equally discoverable, trustworthy, or useful. Google became valuable precisely because an open universe of endpoints still required crawling, indexing, ranking, and learning. Any professional could publish a personal website; that did not eliminate LinkedIn because professional identity, relationships, reputation, and discovery become more valuable when represented in a common graph. Merchants can accept payments directly; that did not eliminate payment networks because shared rules, trust, authorization, acceptance, and interoperability compound across participants.

The same distinction is likely to exist for agents. A public Agent Card can make an agent machine-readable and state what it claims to do. An open protocol can make it callable. A registry can make it searchable. None of those things is automatically an economic history. A network that mediates interactions can know how often an agent has actually been considered, how it responds to different kinds of demand, how it negotiates, whether it delivers, which counterparties trust it, whether disputes occur, and whether the relationship continues afterward.

Much of that information cannot be reconstructed by simply crawling the open web because it is private, dynamic, and path-dependent. It exists because two parties interacted and because the system was present for the interaction. So I do not think the open agent web and agent networks are mutually exclusive. In fact, I expect the protocols at the edges to become increasingly open while the economically valuable graphs built on top of them become increasingly important.

The open protocol makes an agent reachable. The network makes the agent economically legible.

Outcomes close the loop

This is the part of agentic commerce I think is still most underappreciated. Most discovery systems learn from proxies for what the user actually wanted. Search engines observe queries, rankings, clicks, reformulations, and sometimes conversions. Marketplaces observe purchases, ratings, returns, disputes, and repeat behavior. Advertising systems observe impressions, clicks, conversions, downstream revenue, and increasingly incrementality. These are powerful feedback signals, but they generally begin after the user’s underlying intent has already been compressed into a search, listing, campaign objective, or transaction.

An agentic network can potentially observe a much richer sequence because the goal itself can be represented explicitly. The system can know what the buyer was actually trying to accomplish, which constraints and preferences defined success, which counterparties were eligible, which were considered, who responded, what information had to be exchanged, which terms changed through negotiation, who was selected, what happened during execution, whether the result was accepted, whether there was a dispute or revision, whether the buyer returned, whether the parties worked together again, and whether the original goal was ultimately satisfied. The important asset is not any one of those observations. It is the longitudinal chain connecting intent to outcome.

That produces a fundamentally different basis for discovery. A seller can be excellent in aggregate and still be the wrong seller for a particular class of goals. A creator may work unusually well for one category of company and poorly for another. A manufacturer may quote aggressively but miss rush deadlines. A consultant may be expensive but produce outcomes that cause a specific kind of buyer to return repeatedly. Two counterparties may coordinate unusually efficiently because their preferences and working styles have already been learned. These patterns become visible when the system observes repeated economic interactions rather than isolated profiles and clicks.

The feedback loop also makes proactive discovery possible. Once demand persists as state rather than disappearing after a search, the network can react when something changes on the supply side. A supplier gains capacity, a creator becomes available, a price changes, a new software capability launches, or another transaction produces evidence that changes how a counterparty should be ranked. The useful event is no longer that the buyer remembered to search again. It is that an existing intent has become newly satisfiable.

That is when discovery begins to look less like querying a database and more like operating an economy.

Agents create a new economic population

There are roughly eight billion humans, but there is no reason the number of economically active agents should be bounded by the human population. A person may have a durable AI representative. A business may have one. Software can participate directly. Organizations may create specialized agents with authority over particular functions, budgets, or resources. Other agents may exist only temporarily to accomplish one piece of a larger goal.

The more important change is not the precise number of agents but the declining cost of creating economic agency. Humans are expensive participants in markets because we are constrained by time, attention, geography, language, memory, and coordination capacity. Software can maintain vastly more weak relationships, monitor markets continuously, compare alternatives in parallel, negotiate asynchronously, and initiate interactions whose expected value would be too small for a person to justify the effort.

That means agentic commerce should not be measured only by the transactions agents take over from humans. Lowering the cost of discovery, negotiation, coordination, and execution can cause transactions to exist that are currently uneconomic: interactions that are too small, too fragmented, too international, too specialized, or simply too annoying for humans to coordinate. The expansion in economic activity can therefore come from both automation of existing commerce and creation of new commerce.

That is why the long-run opportunity is much larger than “AI shopping.”

The first constraint will not be intelligence

Models will continue improving, and any infrastructure thesis in this market should assume that they do. A product whose value disappears when the next foundation model gets better is positioned on the wrong side of the technological curve. Darwin should become more valuable as models become more capable because better intelligence increases the amount of real-world responsibility an agent can take on.

The bottleneck consequently shifts. The early question was whether a model could understand a person well enough to be useful. Increasingly, the harder question is whether it can reliably get the thing done. A model may perfectly understand a procurement objective and still lack access to supplier identity, private availability, historical performance, authority, negotiated terms, payment, fulfillment state, and verification. The model can know what should happen without possessing the economic state required to make it happen.

That is the transition from intelligence to agency. And once an AI needs to act beyond the boundaries of its own context window and software environment, agency becomes a network problem.

Where Darwin sits

Darwin is building an economic network around that premise. Our underlying belief is that people and businesses will increasingly have persistent AI representations through which they can express goals, expose capabilities, maintain permissions and wallets, build relationships, communicate, negotiate, transact, and accumulate reputation over time.

The important symmetry is that both sides become agentic. A buyer AI does not merely search a static catalog; it can carry the underlying intent, constraints, authority, preferences, and relationship history of the buyer. A seller AI does not merely expose a fixed listing; it can represent dynamic capability, availability, willingness, pricing, permissions, and terms on behalf of the seller. Darwin sits between them and maintains the economic state required to turn one side’s goal into an agreement and ultimately an outcome.

We are starting with human execution because people are the least structured form of economically meaningful supply. Creator work, AI evaluation, research, recruiting, and professional services force the system to deal with dynamic matching, qualification, negotiation, multi-party coordination, execution, verification, and reputation rather than simply optimizing checkout. If the network can reliably represent and transact with humans, then increasingly structured forms of supply become easier to add: digital capabilities, businesses, products, physical inventory, logistics, travel, and infrastructure.

Over time, the goal is for Darwin to span all three execution domains. Some transactions will use open protocols at the edges. Some payments will travel over rails operated by other companies. Some supply will continue to live inside large vertical networks. None of that conflicts with the thesis. Darwin does not need to replace the internet or every underlying market. It needs to become the place where enough economic identities, relationships, intent, and outcome history accumulate that resolving a goal through the network becomes better than reconstructing the same state from scratch every time.

That is the compounding asset.

What would make this thesis wrong?

There is a serious counterargument: sufficiently capable models may be able to use the open web itself as their network. An AI could crawl public Agent Cards, search registries, inspect merchant feeds, call A2A endpoints, negotiate over open protocols, use ACP or UCP for commerce, pay over machine-payment or traditional rails, maintain a private memory of every counterparty, and infer reputation from publicly available evidence. If that architecture can resolve complex goals with comparable trust, latency, liquidity, cost, and outcome quality, the incremental value of a shared economic network becomes much smaller.

I think that is the correct falsification test. The Darwin thesis is not that open protocols will fail. I expect them to succeed. The bet is that economically important relationships contain enough private, changing, relational, and path-dependent state that independent agents continually reconstructing the market from public endpoints will remain inferior to a network that directly observes and maintains that state.

Identity, authority, private capability, availability, negotiated terms, relationships, trust, transaction history, and outcome history all gain value as they accumulate. The open web can make every agent technically reachable. A shared economic graph can make the right agents reliably useful to one another. Darwin only deserves to exist if that distinction compounds.

Intent is only the beginning

The first era of agentic commerce will look familiar because AI will make the economy we already have substantially easier to navigate. Agents will search better, compare more options, purchase structured products with less intervention, invoke paid software autonomously, and gradually receive greater authority to transact on behalf of people and businesses. The existing markets remain visible; agents simply become much better interfaces to them.

The more consequential shift comes when the markets themselves begin organizing around persistent intent. The interface to economic activity moves from browsing supply to describing desired outcomes. Transactions move from objects that humans assemble manually to structures agents can form dynamically. Discovery moves from isolated searches toward continuous matching. Relationships move from fragmented human memory and CRMs into persistent agent state. Human, digital, and physical execution become components of the same goal rather than separate places a user has to navigate.

The internet made the world’s supply accessible. AI is making the world’s demand legible to software. What remains is the system capable of connecting the two, coordinating what has to happen in between, and learning from whether the outcome was actually good.

That is the part of agentic commerce we are building Darwin around.

Part II: Networks.