When AI Agents Pay: Stablecoins as the Wallet of Machines

For most of the internet's history, every payment had a human at the end of it. Someone typed in a card number, confirmed a transfer or clicked a button. That assumption is starting to wobble. Software agents built on large language models can now search, compare, fill in forms and call other services on their own, and some of those services cost money.

If an agent is supposed to buy a dataset, rent an hour of computing power or pay for a single API request, it needs a way to pay that does not require a person to approve every step. This is where parts of the crypto industry see an opening: money that exists as code, moves around the clock and can be controlled by rules written into software.

The idea is plausible, but it is early. Much of what is discussed today exists as prototypes, pilot projects and technical proposals rather than as everyday infrastructure. It is worth looking at both the logic behind it and the questions that remain open.

Why software agents need their own way to pay

An AI agent is a program that pursues a goal over several steps and decides along the way which tools to use. A travel assistant in software form might check flight prices, query a weather service and reserve a hotel. Each of these steps can involve a paid service, and the more capable the agent, the more often it runs into a paywall.

Today, most paid APIs work with user accounts, monthly subscriptions and stored credit cards. That model works when a developer signs up for a handful of services in advance. It works less well when an agent discovers a useful data source on the fly and needs access for exactly one request.

The web anticipated this problem a long time ago. The HTTP standard has reserved status code 402, "Payment Required," for future use, yet it was hardly ever used because there was no simple, native way to pay inside a web request. Several current proposals try to fill exactly that gap.

Micropayments and the problem with fees

The central obstacle is economic. Card payments often carry a fixed fee per transaction in addition to a percentage, which makes a payment worth a fraction of a cent pointless. Bank transfers are cheap in some regions, but they are slow across borders and not designed for thousands of payments per hour.

Stablecoins are tokens on a blockchain whose value is tied to a currency such as the US dollar or the euro, usually backed by reserves held by the issuer. They can be sent at any time of day, and on many networks a transfer costs little. Since the Dencun upgrade in March 2024, fees on Ethereum's layer 2 networks, which bundle transactions and settle them on the main chain, have fallen markedly. That makes small payments more realistic.

This does not mean every agent payment will be a blockchain transaction. Bundling many small charges into one settlement, prepaid balances or payment channels that only record the final result are also options. The point is that machine payments need to be small, fast and cheap at the same time, and existing systems rarely deliver all three.

Programmable money and smart contract wallets

The second argument for crypto is control. A conventional crypto wallet is controlled by a single private key: whoever holds the key can move all of the funds. That is a poor fit for an agent, because a compromised or confused program would have unlimited access.

Smart contract wallets work differently. The wallet itself is a small program on the blockchain, and it can check rules before any payment goes through. On Ethereum, a standard for this approach, known as account abstraction, has been in use since 2023. Typical rules include:

• A daily or weekly spending limit that the agent cannot exceed.
• An allowlist of addresses or services the agent is permitted to pay.
• Session keys that expire after a few hours and only cover specific actions.
• A requirement for human approval above a certain amount.
• The ability for the owner to freeze the wallet or revoke the agent's access at any time.

These limits are the real innovation, more than the currency itself. They turn an open account into something closer to a company expense card with strict rules. The difference is that the rules are enforced by code on the network rather than by a bank's back office.

The question is not whether a machine can pay, but how precisely its owner can define what it is allowed to pay for.

Liability, identity and fraud

Once software can spend money, the hard questions become legal and organizational. If an agent buys the wrong thing, pays a fraudulent service or is tricked into a transfer, who carries the loss? Is it the user who set it up, the developer of the agent, the provider of the underlying model or the service that accepted the money? Current law was not written with autonomous agents in mind, and there is little settled case law.

Fraud is a concrete risk. Language models can be manipulated by hidden instructions in web pages or documents, a technique known as prompt injection. An agent that reads a manipulated product page might be steered toward a payment it should never make. Because confirmed blockchain transfers are usually irreversible, there is no chargeback as there would be with a credit card.

Identity is the third open point. Regulated payment providers must know their customers and follow anti-money laundering rules. In the EU, issuers of stablecoins that count as e-money tokens fall under MiCA, whose stablecoin rules have applied since 30 June 2024. How an agent acting for a verified person or company is identified, and how a service can tell a legitimate agent from a bot used for abuse, is still being worked out.

Where AI and crypto actually meet

For years, AI and crypto were often mentioned in the same breath, as two hype topics without much in common. Agent payments are one of the few areas where the connection is functional rather than rhetorical. Agents need money that software can handle directly, and blockchains offer exactly that: open interfaces, settlement around the clock and rules that anyone can verify.

At the same time, traditional payment companies are not standing still. Card networks and payment providers are working on their own approaches to purchases made by agents, for example virtual cards with tight limits or tokens that only work with a single merchant. Whether agents end up paying with stablecoins, with cards or with a mix of both is an open question.

For businesses, the practical question is less ideological. A data provider or API operator has to decide whether to accept automated payments at all, how to price individual requests and how to protect itself against abuse. These are product decisions that will likely be made long before any grand vision of a machine economy materializes.

What to watch as the idea matures

Anyone following this field should pay attention to a few signals rather than to headlines. Are there open standards that different providers actually implement, or does every company build its own closed system? Do wallet designs make limits and revocation easy for ordinary users, not just for developers? And do lawmakers and regulators clarify who is responsible when an agent makes a mistake?

It also helps to keep the scale in perspective. Most AI agents today still work in supervised settings, with a person reviewing the important steps. Letting software spend money without oversight is a step that companies and households will likely take gradually, starting with small amounts and narrowly defined tasks.

The sober view is that stablecoins and smart contract wallets are a promising tool for a real problem, not a finished solution. The technology for machine payments largely exists. What is still missing are the rules, habits and safeguards that make people comfortable handing a budget to a program.