FINDING DIRECTIONWallets and card tokens are giving software a financial identity built from delegated authority.
This is the economic sequel to Borrowed Authority. But what happens when some of that authority is financial?
Suppose an agent needs a weather forecast, a database query, a few seconds of computing power, or access to a research paper. Instead of encountering a subscription page, creating an account, requesting an API key, and waiting for its owner to enter a credit card, the agent receives a price. It pays. The work continues.
That small transaction sits beneath a much larger idea: software that can purchase the resources it needs, sell services to other software, and manage a budget within limits established by a person or organization.
The industry has given this idea a suitably ambitious name: the agentic economy.
Crypto companies are creating wallets expressly for agents. Card networks are issuing agent-specific payment credentials. Payment processors are building wallets that produce a new virtual card or token for a purchase.
The economy may be embryonic. The account structure is already here.
From Commerce to Economy
“Agentic commerce” usually describes an AI acting as a personal shopper. You ask it to find a nonstop flight arriving before noon or a waterproof jacket under $200. It searches, compares, and perhaps completes the purchase after receiving your approval.
Its agents do not just shop for people. They buy inputs for themselves. They pay for data, inference, storage, software tools, and access to other agents. They might earn money by completing a task, retain a working balance, and spend part of that balance acquiring what they need for the next task.
This is one reason crypto companies find the idea attractive. The banking system normally opens accounts for people and legally recognized organizations. That process includes identity checks, contracts, and a party capable of accepting liability.
At the protocol level, a blockchain wallet asks a narrower technical question: who is authorized to use the signing key? A wallet can give software a financial credential without granting it legal personhood or asking a bank to recognize it as a customer.
Coinbase launched AgentKit in November 2024, giving developers a way to connect agents to wallets and blockchain operations. In May 2025, it introduced x402, a payment protocol based on the long-neglected HTTP status code 402 Payment Required.
The idea behind x402 is simple. A server can respond to a request with a price. The requesting software attaches a stablecoin payment and tries again. Payment becomes part of the same exchange through which software already requests information.
For a person, eliminating a login and checkout page is a convenience. For an autonomous agent, it is the difference between continuing and stopping for help.
Coinbase made the strategy explicit in February 2026 with Agentic Wallets. These wallets let agents hold stablecoins, pay for services, transfer funds, and trade assets. The private keys stay outside the model, while controls can limit spending per transaction and session.
Circle, OKX, and Crossmint have introduced their own agent wallets and payment infrastructure. The details vary, but the pattern is consistent. The wallet combines several functions that were previously separate:
- A financial identity for the agent
- A balance or source of funds
- Permission to propose or initiate transactions
- Limits on amount, frequency, and counterparties
- Protection of the underlying signing key
- Monitoring, screening, and an audit trail
- A way to revoke the agent’s access
What these companies call a wallet is increasingly also an identity credential, a policy engine, and a record of delegated authority.
Cards Without Card Numbers
Crypto is only half of the story. Mastercard, Visa, American Express, and Stripe are building similar systems for agent spending. Each one depends on an existing human or corporate account.
Mastercard Agent Pay issues an agentic token to a registered agent. The agent can use the credential through ordinary card payment fields without receiving the underlying card number. Mastercard has reported authenticated agent transactions in Australia and says all Mastercard issuers in Europe are enabled for Agent Pay at the network level.
Visa Intelligent Commerce replaces card details with a tokenized credential that confirms which agent may act for the cardholder. The person can attach spending limits and conditions. Visa’s Trusted Agent Protocol gives merchants a way to distinguish an approved shopping agent from an anonymous bot carrying payment information.
American Express has introduced an Agentic Commerce Experiences developer kit through which registered agents can receive tokenized Amex credentials subject to cardholder spending and purchase-intent controls.
Stripe has gone even closer to the language of wallets. In April 2026, it launched Link’s wallet for agents. After the owner approves a spending request, an agent can receive a one-time card or Shared Payment Token backed by the cards and bank accounts in the owner’s Link wallet. The agent never sees the original credentials. Stripe’s separate Issuing infrastructure supports single-use virtual cards, spending limits, merchant-category controls, real-time authorization, and immediate revocation.
Both crypto wallets and card tokens give an agent restricted financial authority. A crypto wallet limits what the agent can sign. A card token grants limited purchasing power from a person’s account.
The agent may control a balance or credential. The money and the liability still trace back to a person or organization.
[ FIELD NOTE / CREDENTIAL ]
An agent’s card is a restricted digital credential, often valid only for a particular purchase, amount, merchant, or period.
A Programmable Power of Attorney
An agent wallet is often described as a tiny checking account for a digital worker. A better analogy is a programmable power of attorney.
The agent receives authority that is limited to a task or class of tasks, valid for a period, capped by amount, restricted to approved merchants or services, observable by its principal, and revocable.
Those restrictions define the product and belong in its core architecture.
An unrestricted credit card number would be dangerous agent infrastructure. A language model can be confused, manipulated, or induced to follow instructions embedded in the information it reads. Connecting that model directly to an ordinary payment credential could turn a prompt-injection flaw into a financial breach.
The useful invention is a card that can express which parts of its owner’s authority the AI may borrow.
The International Monetary Fund identifies the underlying architectural problem in its analysis of agentic payments. Agent reasoning is probabilistic. Financial authorization and settlement depend on deterministic rules. A payment must be allowed or denied, executed predictably, and recorded with legal finality.
The IMF’s proposed answer is to separate those functions. Let the agent interpret the owner’s intent and propose an action. Let a policy layer determine whether the action is authorized. Let the payment network execute and record it.
This architecture leaves probabilistic reasoning upstream and keeps financial authority within fixed rules.
[ FIELD NOTE / LIABILITY ]
When an authorized agent buys the wrong thing, the unresolved question is who bears the cost of its mistake.
American Express has announced plans for Agent Purchase Protection covering eligible charges that deviate from a cardholder’s authenticated purchase intent. The protection is not yet generally available. Its existence points to the larger question behind the field note: is an unintended agent purchase fraud, user negligence, defective software, or an ordinary dispute?
Infrastructure Looking for an Economy
The phrase agentic economy invites exaggeration.
Coinbase and others have promoted large transaction counts for x402, but a blockchain transaction does not reveal whether an autonomous agent, a conventional program, or a person ultimately controlled the wallet.
Chainalysis found that x402 activity crossed 100 million transactions on Base while also concluding that much of the early growth came from meme-coin farming rather than agents purchasing useful services.
A July 2026 research preprint offered a sharper warning. The researchers examined approximately 136.7 million x402 settlements on Base. They classified 21.2% as fictitious and another 63.78% as internal settlements within linked clusters. Their conclusion was that raw settlement count measures how easily activity can be manufactured, not how widely an agent economy has been adopted.
The paper has not been peer-reviewed, and its classifications will be debated. Its broader caution is sound: machine-generated transaction volume is particularly easy to mistake for machine-generated demand.
An agent can pay another agent a million times. That does not mean either of them has a customer.
The infrastructure can still be useful even when much of the early activity is speculative. The early web had empty storefronts, inflated traffic, and business models that existed mostly in slide decks. It also had HTTP, URLs, browsers, and payment systems that would eventually support something much larger.
Today’s transaction count mainly shows how cheaply activity can be produced. The stronger signal is that financial companies with very different business models are independently redesigning credentials, authorization, accounts, and merchant acceptance for software actors.
The Nonhuman Customer
For decades, payment systems have been built around a person at the decisive moment. Someone enters a card number, presses a button, signs a receipt, confirms with a fingerprint, or approves a notification. Even automated payments are generally the delayed result of an earlier human instruction.
Agents disturb that model. A person may define the objective and limits without participating in each decision. The agent chooses the merchant, timing, product, and payment method.
[ FIELD NOTE / DECISION ]
The agentic economy shifts purchasing decisions from the person to the agent.
It creates an unfamiliar participant at the interface: software that can compare prices, negotiate, and execute while the legal customer and liability remain elsewhere. Every dollar it spends still belongs to a principal standing somewhere behind it.
With credentials, agents can act like independent customers.
The financial instrument may not be cryptocurrency, a credit card, or a bank account. It may be the restricted token through which an agent borrows just enough of our authority to spend it.