An AI agent completing a task may need to purchase something along the way – access to a dataset, a computation, a service. Nobody is present to enter a card number. That is a new payment problem and several approaches are being tried.
Why conventional rails struggle
Card payments assume a human, a merchant, and a dispute process. Agent payments are small, frequent, automated, and between parties that may both be software. The fees and friction that are tolerable for a human purchase are not for a thousand micro-transactions per hour.
Why crypto rails fit
Programmable. A wallet can be given rules – spend up to this much, only on these things – and the agent operates within them. Transactions settle without intermediaries and without a human confirming each one.
Stablecoins remove the volatility problem, which was the obstacle for every previous machine-payment proposal.
The obvious risks
An agent with a wallet is a target. Prompt injection – tricking the agent into spending – is a documented attack and it is not solved.
Spending limits help. Requiring human confirmation above a threshold helps. Neither is a guarantee.
What is actually deployed
Narrow cases: agents paying for API calls, for data access, for compute. Amounts are small, the counterparties are known, and the rules are tight.
The general case – an agent with discretionary spending power – is being demonstrated and is not being deployed by anyone careful.
Where this goes
Towards a payments layer that assumes machines on both ends. Whether that layer is crypto-based or something built by conventional payment companies is genuinely open, and both are being built.