Over the past year I have watched autonomous AI agents move from novelty demos into real workflows. These agents are now starting to pay each other for services, data, and compute. The payment systems we have used online for decades work well for people buying from other people or businesses. They are not built for agents trading value in tiny bursts, hundreds or thousands of times per day, without human oversight.
Agent-to-agent payments often involve amounts smaller than a cent, happen at very high frequency, and need to settle almost instantly. They also have to be programmable so one agent can pay another automatically as part of a workflow.
In this article I share what I have learned about the payment rails and protocols that make this possible, how the economics play out, and what tradeoffs product teams should keep in mind.
Agent-to-agent payments happen when one autonomous system pays another for a specific service or result.
Some examples I have seen or worked with:
In each of these, the payment is not a monthly subscription or a one-time invoice. It is tied to a single unit of value and happens in seconds.
Stripe with Model Context Protocol (MCP) can technically allow an agent to initiate payments. But when you run the numbers and look at how agents work, it is not always a fit.
These factors push most products toward bundling or subscriptions, even when micro-metering would be more precise and fair.