Sooner or later the agent you built will need to pay for something. It already runs your support queue, or your research, or a piece of your operations, and the next thing it needs is a data feed that charges per call, or an hour of another agent’s work, or a domain and hosting for the page it built. The card on file is yours. The bank account is yours. Neither was designed to be handed to software that decides at 3 a.m. which purchase to make.
The first instinct is to give the agent an API key with a budget and hope. There are more native options now, built for software that pays rather than for a person holding a card, and they differ in how much stands behind them: some are live and measured, some are only announced.
The word token needs a definition first, because in 2026 it means two different things. In AI, a token is the unit a model charges by: the model reads and writes text in small pieces, and the bill counts the pieces. On a blockchain, a shared ledger kept in copies on many computers, a token, also called a crypto token, is a digital coin that a project creates, gives a name, and lets people buy, hold, and sell. This article uses the second meaning. A stablecoin such as USDC is a token in that sense, one that stands for a dollar. When a project launches a token of its own, a native token in the language of token design, it has created a new coin, named for the project or its agent, and opened it to trading. A wave of projects did that for their agents in 2024 and 2025, and what came of it tells you which road holds.
What an agent can pay with today
A wallet with limits.
The gap between “an agent that recommends” and “an agent that acts” was money it could hold in its own name without a human approving each step. Wallet infrastructure built for software has largely closed it. Coinbase’s Agentic Wallets, launched February 11, 2026, give an agent a wallet with a session cap and a per-transaction limit; the private keys, the secrets that authorize spending, “remain in secure Coinbase infrastructure, never exposed to the agent’s prompt or LLM.” Privy’s agent wallets take policies that “cap transfer values per transaction or over a time window, allowlist recipient addresses and smart contracts,” and cap the value of any single request. On August 18, 2026, AWS made AgentCore Payments generally available, with the limits enforced “at the infrastructure layer” rather than in the model.
The spending rule lives below the model, where a prompt cannot change it. In practice that is a policy like: no more than $20 a day, no single payment over $1, and only to these three services. A bounded wallet is the first thing to give an agent that has to make purchases.
A dollar that moves at the speed of a request.
Per-call pricing needs a payment that settles in seconds for a fraction of a cent. A stablecoin, a token that stands for one dollar held in reserve by its issuer, does that when it is sent over HTTP. The protocol behind much of it is x402: a server answers a request with “payment required,” the agent’s wallet pays, the server delivers. In July 2026 x402 moved under a Linux Foundation body whose 40 members include AWS, American Express, Circle, Google, Mastercard, Shopify, Stripe, and Visa. Membership signals intent; the use is measured elsewhere.
The rail is not limited to one coin; x402’s own site says “stablecoin payments are the primary use case” and that it is extensible to other payment methods. What settles on it is a tokenized dollar. Circle reported that 99.3% of x402 payment volume settled in USDC in the second quarter of 2026.
How much of that volume is real commerce is a fair question, and the answer so far is a careful one. A population-scale study of x402 on Base, a network Coinbase built on top of Ethereum, counted 136,708,672 settlements worth $44.1 million over 280 days. It classified 21.20% of them as fictitious and 63.78% as internal to linked operator clusters, sets of wallets the study traced to one operator paying itself. Genuine outside value, on its bounds, sits between $187,861 and $20.26 million; the floor is the part the study could not attribute to any operator’s own cluster, which may be independent or unresolved. The authors’ line is that “settlement count measures manufacturability, not adoption.”
The same study looked at the sellers. The 25,163 advertised services collapse to 811 distinct recipients, and among the 624 that ever settled a payment the median seller earned $3.96 over its lifetime. When the service routing most of those payments stopped doing it for free, on January 1, 2026, volume registered its largest sustained fall, “volume behaving as if priced at the subsidy.” The rail runs, but the sellers on it are few, and how much outside commerce reaches them is unresolved.
The card networks came to the agents.
The other half of the story is that the conventional rails did not get displaced; they absorbed the new customer. On March 27, 2026, Mastercard completed what it called its first live, authenticated agentic transaction, in Hong Kong. An agent booked a ride to the airport and paid with a tokenized card credential, a substitute card number scoped to that agent rather than the card itself, and the cardholder confirmed it with a passkey, the same device tap used to sign in without a password. In June Mastercard launched Agent Pay for Machines, which “supports reliable, guaranteed multi-rail settlement across cards, accounts and stablecoins,” some payments “only fractions of a cent,” with more than 30 partners named at launch. Visa’s answer, in pilot since April 2026, is an adapter: Intelligent Commerce Connect lets a merchant accept payments “initiated via major agent protocols including: Trusted Agent Protocol, Machine Payments Protocol (MPP), Agentic Commerce Protocol (ACP), and Universal Commerce Protocol (UCP),” four competing standards behind one on-ramp built to work with any of them, so the merchant does not have to pick a winner. Launch is announcement, not volume. Even so, the pattern is clear enough to plan on. When the agent buys something for a person, a scoped card credential is the natural rail, because refunds, disputes, and the merchant relationship already live there. When it buys a machine service, a stablecoin is.
Hiring another agent.
At some point the agent will want to buy another agent’s work, and one design that fits is escrow: funds locked to a job, released on completion. Virtuals, a platform where agents launch with tokens of their own, runs an Agent Commerce Protocol for hiring one agent from another; it has denominated its jobs in USDC since August 2025, even though new agents on the platform still launch with the VIRTUAL token. Olas runs a marketplace where one agent pays another for a bounded task; its counter showed 14,470,430 agent-to-agent transactions against $108,339 of turnover on August 27, 2026, well under a cent each. Both are working infrastructure; the figures they publish do not separate independent-buyer volume. In the Virtuals case the money the agent hires with is, again, a dollar. Each agent on the Olas marketplace, a mech in Olas’s term, declares its own payment asset in its contract: the blockchain’s own coin, OLAS, or USDC.
Those four have shipped as of September 2026: a bounded wallet, a stablecoin paid per request, card rails with an agent credential, and escrow for agent-to-agent work. You can pick a starting rail now; the rails work without the need for a native token.
How we got here: the agent-token wave
Thousands of agents got a token first
Between late 2024 and early 2025, thousands of projects took the other road: the agent got a token at launch, and the token was the story. Coinbase’s institutional research team estimated in February 2025 that “agentic AI” crypto assets peaked above $20 billion in market capitalization, the value of all the tokens together at the going price, in early January and ended the month closer to $8 billion, with “unclear long-term utility” for agent tokens. On Virtuals, which that note called the leading AI launch platform on Base, nearly 16,000 agent tokens had been launched by then, and about 2% of them could be bought and sold on an open market.
A study published in May 2026 followed what happened next for one slice of that wave: agents built to invest their holders’ money. Out of more than 1,900 crypto projects tagged as AI, most of them not agents at all, the researchers filtered to the AI agents built to invest and curated ten and measured them. Combined valuations had passed $3 billion. The agent tokens in that study fell 93% on average from their all-time highs, in a range of 88% to 99%. The projects’ own treasuries, the holdings each team controls, kept about $34.3 million in paper gains, up in value but not sold. Token holders, 925,323 of them, ended $191.7 million underwater on the same paper measure, down from about $2.4 billion of gains at the peak, and the top 1% of profitable wallets took 81.4% of all the gains.
Two agents’ tokens were valued at more than 10,000 times the assets they managed, more than $10,000 in token value per dollar of assets held, where established finance protocols that run on the blockchain, such as Uniswap, an exchange, and Aave, a lending protocol, sit below 1×. When the researchers interviewed the teams behind the two main frameworks, ElizaOS and Virtuals Protocol, the answer on whether the agents could invest on their own was direct: “LLMs cannot trade well” without human insight, and “the vast majority” of launched agents were “basic API integrations.”
One thing the tokens left behind was an audit trail. The four projects in the study that raised money through a public token launch left a public treasury and an on-chain footprint the researchers could examine; the four funded through private user wallets left “no public treasury to audit.”
The wave continues at a smaller scale. Virtuals’ own documentation, current as of August 14, 2026, still describes the launch mechanism: “Trading opens when the agent is created.” The category holds about $3.2 billion in market value as of September 2026, against the more than $20 billion of January 2025, and VIRTUAL, the platform’s own token, sits 87.8% below its high of January 1, 2025. Launches continue; the market value around them is a fraction of the peak.
The market’s own research desk now points at the rails
Fourteen months after that peak, in March 2026, Coinbase’s research desk published where it sees the opportunity now, and it leads with the rails. Coinbase’s Picks-and-Shovels of the AI Agent Economy, March 31, 2026: “We view the real AI agent opportunity in crypto as infrastructure: the wallets, payment rails, and settlement layers that let autonomous software transact at scale.” The note calls USDC “the reserve asset of the machine economy,” and expects that “the first large, non-speculative value pools are more likely to appear in systems that settle and route machine commerce than in the speculative app-layer tokens.” In that note’s terms, the bounded wallet, the stablecoin paid per request, and the escrow for hired work are the infrastructure, and the agent tokens are the app layer.
The note says where the value went. The W5H framework for token design, which I published in January 2023, sorts the same cases by what the token is for, and the sorting explains why.
Why would an agent need a token in the first place?
In January 2023 I published the W5H framework for token design, Why, When, What, Where, Who, and How, and it starts from one question: why would a project need a token in the first place? A project there means any team building a product on a blockchain, and the answer sorted the projects that use or issue a token into three types by what the token is for. Type A projects process tokens that already exist, without the need to create their own. Uniswap, the exchange, started without a token; when UNI came later it was for governance, voting on changes to the protocol, and not required by the core exchange service. An agent that pays for data, compute, and other agents’ work with a stablecoin is a Type A project, a customer of tokens rather than an issuer.
Type B projects create a token because the token is the product: an asset represented on a ledger. USDC is the example, a fiat dollar tokenized for spending in the crypto world and beyond. A stablecoin cannot exist without its token, so for this type the question answers itself, and that is also what the money on the machine rails is: a token whose reason to exist was settled before any agent needed it.
Type C projects create a token to coordinate many independent parties at scale, and the test is whether that incentive is essential to the product. Bitcoin directs the value its network creates to its miners, and Ethereum pays its stakers, who lock up ETH as a deposit, to secure the chain; those are the computers that keep each network running, spread across operators no one company controls, and without the token the network stops. Where a token serves governance, the framework’s answer was “maybe,” since many products start without one and add it when they are ready to hand control to their users. Olas is a Type C case in the agent world today: anyone can lock up OLAS as a deposit to run an agent, and rewards are paid on the agent’s measured activity, not on the lockup alone.
Sorted by the W5H types, the agents in the May 2026 study that launched a token, four of the ten, were Type A projects, software that would consume data, compute, and other services, carrying Type C tokens. The token was not the product, because the product was an agent, and the study does not show the token coordinating independent parties the way BTC pays its miners. The rationale that remains is one the framework lists as a possible answer: tokens “may help raise capital for the project,” and let early community members “invest in the project and gain upside.” The second part of the series added a timing rule for the rest, to launch a token “when its utility becomes indispensable for the crypto project.” Peak assets under management arrived after the token’s peak price, a sequence the study found repeatedly and reads as “a reflexive process” in which “capital inflows into agent treasuries follow only once speculative activity has already peaked.” The fifth part of the series had described, in terms of token velocity, why a token used only to pay for things keeps little of the value: it turns over fast, and each turn keeps almost none, so “a crypto project which creates a lot of economic value does not necessarily accrue value to its tokens.”
The study itself stops short of a verdict on the model, reading what it measured as “signs of an immature market structure rather than as evidence that the underlying platform model is inherently flawed.” Still, the tokened projects in the study carried more than $3 billion in combined valuation at the peak, the question ran in public for a year, and the rail underneath kept working through it. USDC in circulation grew 19% year over year to $73.3 billion by mid-2026, and Coinbase’s and AWS’s agent wallets shipped after the January 2025 peak. The rails that shipped use a token that already existed: the dollar, tokenized.
Still open: identity, accountability, and delivery
The rails settle payment, and they will keep developing. Three things they do not settle yet are open, and they are the ones to watch.
The first is identity. A wallet address proves that a key signed; it does not say who is accountable for what the key did. The a16z crypto team put it in January 2026 as “the bottleneck for the agent economy is shifting from intelligence to identity”. One open standard, ERC-8004, gives agents an identity, a reputation record, and a way for others to check their work, recorded on the blockchain, without any token of its own; the standard’s own text says “payments are orthogonal to this protocol,” meaning separate from it. A standard with no token of its own is the answer the why-token question points to. But registration is not trust. A study of 173,441 registrations found that only 3%, 4%, and 15% of agents on Ethereum, BNB Chain, and Base had a valid registration record, the file that says where the agent can be reached, with at least one declared address for its service. It also found that 73.5%, 59.2%, and 90.6% of reviewers showed coordinated Sybil behavior, one operator posing as many reviewers. A registration count says more about how cheap registration is than about how many agents work: posting feedback costs $0.0027 on Base against a median of $0.70 in payments an agent’s wallet had received, the value a fake review can move.
The researchers name the cause: submitting feedback costs only the blockchain’s transaction fee, which the network calls gas. One fix they name is to make influence cost something, “a stake that is slashed after a successful dispute,” meaning the stake is taken from the operator who posted it. In W5H terms, that is a Type C mechanism. Read that way, if a why-token case forms in the agent economy, identity is a likelier place for it than payment, and no such case appears in the studies here.
The second is accountability. In July 2026 HM Treasury opened a consultation on modernizing the UK’s payment rules and asked: “do provisions relating to authentication and consent of payments transactions, and liability for unauthorised payment transactions, need updating?” The UK’s Financial Services AI Adoption Plan, also dated July 14, 2026, names “Legal & Liability Frameworks: Defining clear legal constructs and dispute mechanisms to unambiguously assign accountability when autonomous agents transact” as one of three pillars of an agentic-payments trust framework, beside Know Your Agent protocols and authentication standards. The IMF’s note on agentic AI in payments lists “legal uncertainty” among the risks it reviews. The consultation closes on October 6, 2026.
The third is delivery. A payment hash, the transaction record a blockchain hands back when money moves, proves that a payment happened. It does not prove that what was paid for was delivered: that the data was correct, the model’s answer was usable, the service ran, or, for a physical purchase, the goods shipped. The rails settle money; whether what was paid for arrived is outside their scope. Coinbase’s note calls this the integrity layer, “proving that agents actually executed the work they claim to have done,” and puts it above the rails in the stack. That is the layer where an agent economy becomes an economy rather than a payment network.
The Takeaway
Today an AI agent can pay with a bounded wallet, a stablecoin paid per request, a card credential scoped to the agent, and escrow for work bought from another agent, without the need for a native token. Between late 2024 and early 2025, thousands of agent projects launched a native token instead. In the one study that measured ten representative investment agents from that wave, the agent tokens fell 93% on average from their highs, while the rails underneath kept working. I concur with Coinbase Institutional’s view that the durable value is in the rails and that the long-term utility of agent tokens was unclear. The W5H why-token question I raised in 2023 still applies to answering that question. As the rails keep developing, the rest of the things yet to settle, such as who the agent is, who answers for what it did, and whether what it paid for was delivered, are worth monitoring closely as well.
