Why AI agents pay in fractions of a cent

Sample article · roughly 900 words · fixed content for testing

The unit of work is getting smaller

Human software buyers think in months. They sign a contract, get a seat, and pay a predictable monthly amount whether they use the product twice or two thousand times. Autonomous agents do not work that way. An agent that researches a question may touch forty web pages, two search indexes and one verification service in under a minute, then never call any of them again. Charging that workload a monthly subscription is absurd on both sides: the seller either overcharges for a burst or undercharges for a flood.

Nanopayments solve the mismatch by pricing the smallest useful unit of work — one page read, one claim checked, one address scored — and settling it immediately. The price of a single call sits somewhere between half a cent and five cents, low enough that an agent can spend without a human approving each line item, high enough that a service with real volume earns a living.

Why card rails cannot do this

Card networks carry a fixed cost per authorization, typically around thirty cents plus a percentage. A one-cent charge on that rail loses money by a factor of thirty before anything else happens. Stablecoin transfers on a low-cost chain invert the economics: the network fee is a rounding error against the payment, so the payment can be as small as the work it pays for.

The second problem card rails cannot solve is identity. Cards assume a person with a billing address who can be charged back. An agent has a wallet and a signature, no address, and no dispute department to call. Payment therefore has to be final, cryptographic and verifiable by the seller alone.

The 402 handshake

HTTP reserved status code 402 for payment in 1997 and left it unused for a quarter of a century. The x402 pattern finally gives it a job. The buyer calls an endpoint with no payment. The server answers 402 and includes machine-readable requirements: which asset, which network, which receiving address, and how many atomic units. The buyer signs a transfer authorization for exactly that amount and repeats the request with the signature in a header. The server verifies the authorization, settles it, and only then does the work.

Nothing in that loop requires a human. There is no signup form, no API key to provision, no invoice to reconcile at month end. A newly spawned agent with a funded wallet can discover a service and become a paying customer within one round trip.

What sells in volume

The services that earn on this rail are boring and constant. Reading a web page cleanly is the clearest example: almost every research, monitoring, shopping and summarization agent needs it, most of them are bad at it, and the output is easy to judge. Address and contract safety checks follow the same shape. So does source verification, where an agent wants to know whether a sentence it is about to repeat is actually supported by the page it came from.

Novelty is not the moat. Reliability is. An agent that gets a clean answer in under two seconds, every time, at a stable price, will keep calling the same endpoint for as long as it exists. Latency, uptime and predictable output schemas matter more than clever features, because the customer is a program that will silently switch providers the first time a response fails to parse.

Designing for a machine customer

A machine customer reads documentation exactly once, at discovery time, and expects it in JSON. It wants error codes rather than error prose. It wants the price expressed in atomic units, not in a marketing string. It wants the same field present on every response, even when empty, so its parser never branches. Meet those expectations and integration takes a minute; miss them and no amount of landing-page copy will rescue the sale.

The last design question is caching. Two agents asking for the same popular page within a few minutes should both be served, both be charged, and only cost the seller one fetch. That gap between price and marginal cost is where a nanopayment service actually becomes a business.