The Agent Economy: Who Owns the Rails AI Runs On

TL;DR

• The agent economy is far bigger than payments. Global commerce runs on outdated rails. From supply chains to contracting to treasury and settlement, agents can make numerous industries significantly more efficient and truly global.
• The rails are being set now, and the stakes are a generational wager. The standard infrastructure chosen in this cycle decides whether users own their agents, data, and power of choice, or a few centralized platforms. It’s difficult to change infrastructure once it hardens.
• NEAR is building an open, integrated stack for the agent economy. NEAR intends to bring together identity, liquidity, private inference, confidential execution, settlement, governance, and economics as a single system.
• This integrated stack is what makes NEAR the foundation for its "AI money" thesis. The NEAR token can support network security, coordination, settlement, and ecosystem economics as agent-driven activity scales.

The race to set the rails

The incumbent payment giants already recognize that the agentic shift is here. In a matter of months, agents went from a feature these companies supported to a market they are building for. Stripe and Paradigm launched Tempo, alongside an open Machine Payments Protocol for agents to pay for services. Visa, Mastercard, and JPMorgan each published an agentic commerce framework. And a16z has mapped thestructural gaps that still prevent agents from acting as full economic participants.

The convergence reflects a shared recognition that software transacting on its own behalf is inheriting a financial system that was not built for its capabilities. It also reflects something more consequential: the default infrastructure is being set now. Early internet architecture was chosen once and lived with for decades. The infrastructure that governs how agents move value and prove their work will likely harden the same way, and whoever sets that standard will capture the value that flows across it.

The largest enterprise software companies are now publishing AI sovereignty doctrines of their own, framing control over data and model weights as the precondition for truly owning your business. Even the incumbents are concluding that, for the enterprise, whoever owns the infrastructure owns the choices that run on it.

This is the defining question of the agent economy: not whether agents will transact, but what infrastructure they will transact on, who controls it, and what becomes of the value created by the people whose capital, data, intelligence, and intent those agents act on.

AI is about to become one of the largest creators of economic activity in human history. As Delphi Ventures recently forecasted, “The Agentic Economy is imminent. The internet is rapidly being remade for agents.” NEAR’s firm conviction is that this economy should run on an open standard and neutral rails.

What the agent economy actually means

The terms in this space are often used interchangeably, but they describe distinct concepts.

The agent economy is the full system: software agents acting as economic participants (discovering counterparties, negotiating, contracting, and settling) on behalf of the people and businesses they represent. The agent economy is a market category, not a single vertical.

Agentic commerce is the broader market trend that analysts are sizing, the consumer-and-merchant layer where an assistant books a flight or reorders supplies.

Agentic finance is the vertical where the proof is clearest today: agents and users moving capital, settling trades, and paying for services.

Agentic payments are a small subset of these terms, the settlement mechanics underneath. Swapping an API key for a crypto payment is a narrow advance. The harder problem, and the one that defines the category, is the full commercial flow: an agent finding a counterparty, agreeing terms, executing, resolving disputes, and settling, with reduced or no human involvement.

Most of the current attention collapses the agent economy down to the payment layer, the one layer existing rails can already serve. The structural opportunity is the coordination layer at the center of it all.

Nearly every commercial relationship that exists today, from a supply chain reordering stock to a business paying its contractors to a treasury rebalancing, is a transaction waiting for an agent to orchestrate it. These relationships run today at human speed, gated by approvals, business hours, analog processes, and reconciliation. Handed to agents, they can run continuously, at machine speed, and the agents that run them can be specialized and more efficient than the manual processes they replace.

The agent economy is not simply a sped-up version of the commerce that exists. As agents take over more of the flow—first payments, then procurement, contracting, and coordination—they do two things at once. They compress existing value chains, collapsing steps that once required teams and intermediaries. And they open new ones: markets in machine-negotiated services, agent-to-agent contracting, and continuous micro-settlement that no human-paced process would attempt. In this paradigm, a new layer of economic activity emerges, interwoven with today's fiat commerce and crypto rails but extending both into categories that are native to agents.

The shift underway is not the digital-to-agentic echo of the move from physical to online commerce. It is the emergence of a whole new category of value.

Who owns the agent economy

AI is on track to become the dominant interface to computing. A growing share of search, commerce, and workflows will route through agents acting on behalf of humans. That makes ownership the decisive question, because the entities that own the models, the interfaces, and the data can shape the decisions that flow through them. The trajectory of the current status quo is toward a handful of companies owning the models, the interfaces, and by extension the choices of many people and businesses.

The control points already exist, and they sit at every layer of the stack. At the model layer, access is a permission that can be repriced or withdrawn: in early 2026, a leading AI lab cut off third-party agent tools from reaching its models through consumer subscriptions, moving that usage to a metered tier withunder a day's notice, a business decision that disrupted workflows built around subscription-based access overnight. At the terms layer, many model-provider termsreserve the right to suspend or terminate access at the provider's discretion, and to use submitted data to train the provider's own systems unless the user actively opts out. At the hardware layer, the chips that run frontier AI are themselves subject to control: the most capable processors have beenrestricted, re-permitted, and restricted again as policy shifted, with access revolving around decisions made far above the businesses that depend on them.

Together these cases point to a larger structure: when the models, the interfaces, the terms, and even the silicon are closed source and owned by a few centralized parties, access to all of it becomes opaque and conditional. The agent acting on your behalf runs at the discretion of whoever controls the layer beneath it.

The other path is a system where users retain ownership as AI becomes the primary way they interface with  the internet, and therefore the global economy. Keeping ownership takes more than open-weights models or self-hosting. It requires guarantees built into the infrastructure itself: privacy, so your data stays yours; verifiability, so you can prove what ran, on what data, with which model; openness, so markets remain broadly accessible and not controlled by a small group of actors; and alignment, so the model is optimized for your success.

User-owned AI is the whole stack, from inference to settlement, not a license on a model.

What agents need, layer by layer

For an agent to act as a real economic participant, it needs a specific set of capabilities at once. The structural gaps are well known. a16z'sanalysis names several of them, and the full set runs wider still.

Seven capabilities let an agent operate autonomously:

• Identity and permissions. An agent needs a durable identity it controls, with scoped authority over what it can spend, sign, and access, and the ability to carry reputation across interactions rather than starting from zero each time.
• Settlement at machine speed and micro scale. Agents transact continuously, in amounts and frequencies no human-paced rail was built for. Clearing has to be fast, final, and cheap enough that a fraction-of-a-cent payment still makes sense.
• Private inference. The model an agent reasons with sees its most sensitive inputs. Those inputs, and the identity behind them, have to stay private from the model provider, not just encrypted in transit, or every decision an agent makes is visible to whoever runs the model.
• Confidential execution. Beyond inference, anything sensitive an agent touches—a trade, a credential, a private dataset—has to run without broadcasting it to competitors or counterparties.
• Verifiability. When agents act faster than humans can review, trust has to come from proofs rather than oversight: a way to confirm what ran, on what data, with which model, and that the result is what it claims to be.
• Universal liquidity. Value is currently fragmented across traditional finance and crypto markets across dozens of chains, and across venues that don't share order flow. An agent needs to reach fragmented liquidity through a unified interface, without managing bridges, gas tokens, or a separate integration for every chain and exchange.
• Discovery and coordination. To form an economy rather than a set of isolated bots, agents have to find each other, compose services, and pay one another for work.

Those seven capabilities empower an agent to operate. Two more determine whether the larger economy they act within can persist:

• Transparent governance. Infrastructure that coordinates real economic value needs an explicit, accountable, and open way to set its rules: how upgrades are decided, how parameters change, how the system adapts as usage scales. The more agents depend on a system, the more its governance has to be legible and contestable.
• Sustainable economics. The system has to capture value from its own usage and route it back into security and longevity, rather than relying on token emissions or subsidies that eventually run out. Activity should strengthen the network that carries it, and the participants who create that value, the users, validators, and other participants who secure and/or use the network, should participate in the system’s long-term sustainability rather than watching value accrue only to a platform above them.

No single product delivers all these capabilities, and the operational and structural layers are usually owned by different, uncoordinated parties: one company for the model, another for settlement, a foundation for governance, none of them aligned. The opportunity is in coordinating these layers into a coherent, open stack, and that is the gap NEAR is building to fill.

Value movement in the agent economy: NEAR’s AI money thesis

In an economy where agents are the primary actors, the asset that anchors value, settles work, and secures trust forms a new monetary category. NEAR's term for it is AI money. This monetary thesis is built backward from how value behaves once agents transact at scale.

Why agents will live onchain

Agents will reason on models and call services all over the internet, but the economy they participate in (where they hold value, settle work, and coordinate with each other) has to run on open and neutral rails. The incumbents can clear an agentic payment, but what they are not structured to provide is the full set of capabilities agents need at once: settlement that crosses chains and asset types without having to manage bridges and fragmented liquidity, execution that is confidential from public view, trust that is programmable rather than permissioned, and economics that favor the participants rather than the operator of a closed network.

Each of those requirements cuts against how traditional rails are built. A closed network can add a payment endpoint for agents but it cannot become open, neutral infrastructure. That is what makes the coordination layer for the agent economy an onchain question.

Existing monetary constructs in crypto

Crypto has reached for monetary status before. Bitcoin's "sound money" rests on a fixed supply: credible scarcity (3-4% inflation in practice), but no native yield and no productive role. Ethereum's "ultra sound money" thesis went further, burning a portion of fees so that network usage would shrink supply. The mechanism was real, but it is coupled to congestion on Ethereum's base layer, and as activity moved to L2s after the Dencun upgrade, ETH drifted back to mild net inflation of roughly 0.23% a year.

The lesson with these designs is that a token captures value only where it is genuinely needed to pay for something. Ethereum's burn is wired to generic blockspace demand on its base layer, which its own scaling moves elsewhere. NEAR’s core applications are designed to connect product usage, application-level fees, and protocol-level economics more directly to the NEAR token and to the capabilities agentic applications require. As those applications generate revenue, that revenue from those can support network-aligned mechanisms such as buybacks and other activities that may remove NEAR tokens from circulation. The objective is for the network to become stronger as usage grows, including as agent-driven activity scales, with revenue supporting security, sustainability, and utility rather than relying only on emissions or external subsidies.

How agents complicate the three classical functions of money

Each of the three classical functions exists to solve a specific friction that comes from humans being the ones transacting:

1. Store of value solves delayed consumption: you earn now, spend later, and need something that survives the gap.
2. Medium of exchange solves the double-coincidence-of-wants problem: barter fails when what you have isn't what I want, so you need a common intermediate asset both sides will accept.
3. Unit of account solves a cognitive limit: nobody can hold a million pairwise exchange rates in their head, so you need one yardstick to price everything against.

However, none of those frictions describe an agent.

Agents don't delay consumption the way a human does. To the extent anything in the system needs to survive across time, that's a property the human or treasury capitalizing the agent needs, not one the agent needs mid-loop. Agents don't have a matching problem either. They interact with posted-price services, not barter counterparties, so they don't need a common intermediate to solve a coincidence-of-wants problem. What they need is settlement with finality, at machine speed and machine granularity. And agents don't have the cognitive limit that makes a single price yardstick valuable. They compute cross-rates instantly, so the unit of account, in its literal sense of pricing convenience, is close to meaningless to them.

What agents actually need is a way to price counterparty risk with limited practical reliance on traditional legal recourse for each interaction. That's a mechanism-design problem, not a pricing-convenience problem, and the classical monetary framework doesn't have a category for it. It often gets characterized as "unit of account" because that's the closest mental model, not because it's the same function.

A new monetary framework for agents

When asking what a computational, continuously-active, stranger-to-stranger economy structurally needs, the three monetary functions survive, but the content inside them changes.

Store of value becomes network sustainability: not an inert, scarce asset that holds still, but an asset whose role is tied to the security, reliability, and economic coordination of the network it supports.. This is what NEAR's fee switch and staking mechanics are designed to support. Calling it "store of value" doesn't match a mechanism built to connect usage to network sustainability.

Medium of exchange becomes a settlement asset: not the thing both sides prefer to hold, but the asset in which value clears with finality and no counterparty risk, cheap enough to settle sub-cent, machine-frequency transactions, and required at the wholesale layer even when it's invisible at the retail layer. This is a clearing property, not a preference property. Agents don't need to want NEAR; they need transactions to be supported by network-level infrastructure in which the NEAR token plays a role. That distinction is what survives abstraction. Agents may pay in stablecoins at the interface, the way crypto UX today lets a user pay gas in one token while a relayer fronts another and gets reimbursed underneath. That retail layer doesn't threaten the settlement asset's role, because whoever runs the abstraction still has to hold and continuously replenish NEAR tokens to keep clearing what it fronts. Most oil buyers never hold a barrel; the settlement currency sits one layer back, where the underlying unit changes hands. Demand there is real even when no agent touches the token directly.

Unit of account becomes a bonding standard: not a price yardstick, but a slashable, economically-painful commitment a stranger can post with limited reliance on courts or reputation history. You can't slash someone else's stablecoin. Issuance and enforcement have to sit within the same protocol.

An emerging function with no classical analog at all is a metering unit: fine-grained enough to price machine-native resource consumption, inference by the token, compute by the call, directly. That's closer to a utility commodity, priced like electricity by the kWh, than to money in any sense of the classical framework.

The settlement chip (stablecoins) vs. the coordination asset (NEAR token)

This redefined monetary framework is also what separates AI money from the stablecoin it will transact alongside. A stablecoin is a settlement instrument. It is useful for moving a dollar, and bounded to that. NEAR's role is the coordination asset of the system the settlement happens inside: the asset that secures the network, prices its trust, and supports its economic coordination. The agent economy will use stablecoins to settle, the same way it uses many chains to execute. That is exactly why the coordination asset beneath them remains important: it is connected to the security, fees, incentives, and application-level economics of the system, not only to any single user-facing transaction.

Sizing the agent economy

The market slice analysts can already size is agentic commerce, and the numbers are large. Morgan Stanley estimates that agents shopping and transacting on people's behalf in e-commerce alone could drive on the order of$385 billion in annual volume. McKinsey and Juniper put the broaderagentic commerce opportunity in a similar range, with real scale arriving toward the end of the decade.

Those figures measure consumer-and-merchant commerce, the one part of the agent economy that maps onto a market analysts already know how to size. The full agent economy, as defined earlier, is wider: agent-to-agent contracting, machine-native services, and the continuous settlement between agents that no current model sizes because it does not yet exist at scale.

NEAR's addressable surface is not any single vertical inside that, but the settlement and coordination layer beneath all of them: not a chain competing on throughput, but the layer that routes intent to whichever chain executes it best, through one integration.

These forecasts put the category's real scale toward 2030. NEAR is building the coordination layer now so that the default of an open standard is in place as the market matures.

The wager: open rails or closed ones

The agent economy is being built. The open question is whether it runs on infrastructure designed for the people who own the capital, data, intelligence, and intent that agents act on, or on infrastructure that reproduces the extraction patterns of the current internet at machine speed.

Every agent that earns does so because a human supplied the capital, the data, the intelligence, or the intent behind it. On closed rails, the value that agent creates accrues to the platform. On open rails, it can flow back to the participants who create, use, and secure the system. This is not only a question of values. It is what the monetary design decides. When usage, fees, and governance are connected to the network itself, value can support the people and infrastructure that make the system work rather than pooling only at the platform layer. That is the role NEAR’s AI money thesis is designed to play.

NEAR's thesis is that the open path can also be the one that performs best, that the same integrated stack which keeps users in control of their data, their agents, and their value is also the most capable way to run an agent economy.

The infrastructure is being chosen now. The standard, once set, will be lived with for a long time.