Where are agent payments going?

Coinbase publishes x402, an open standard that finally puts HTTP status 402 to work. For five months almost nothing happens.

PING lets anyone mint a token by paying 1 USDC over x402. Weekly transactions jump more than 10,000% in a single week.

Nearly 69 million transactions in a month. Much of it is a single fleet paying itself, plus operators farming for expected rewards.

The largest facilitator starts charging a tenth of a cent per settlement on 1 January. The curve registers its biggest sustained fall, and flattens.

x402scan counts 236 million x402 transactions and $56.6M settled to 30 October. Other trackers land either side: 195.8M and $42.3M at agenteconomy, 279.7M and $58.3M on Token Terminal's undated page. None of them says it removes wash trades. Visa and Artemis do, along with tests, and count 109.6M worth about $15M, but only through 21 April.

Between September 2025 and June 2026, agents made 136.7 million x402 payments on Base, worth $44 million. It is the number quoted whenever someone says the agent economy has arrived.

A July study by Ling, Wang, Zhou and Wu traced every one of them. I rebuilt its results square by square.

Wallets sending money to themselves, or around a closed loop and back. Onchain that is a valid payment, but nothing is bought: it is like moving €10 from your current account to your savings and back.

One fleet alone produced 27 million payments in November, then stopped within two seconds.

Another 64% went between wallets funded by the same operator. One cluster took in $415,765 and swept 98% of it straight back to its funder.

The remaining 15% has no clear owner. The study's behavioural model flags nine in ten of those as manufactured too: flat around the clock, no daily rhythm of real users.

By value, more than half of all the money was a wallet paying itself.

$187,861. Less than half of one square. The study calls settlement count "a Goodhart metric": it measures the incentive to transact, not adoption.

Every year about $2 quadrillion moves between banks, businesses and people. Visa and Mastercard, the rails most of us think of as "payments", carry about 1% of it.

Strip out trading and bots and stablecoins settled about $390B of real payments in 2025: salaries, supplier invoices, remittances. Next to the card networks, a visible sliver.

Agent payments over x402 settled $1.5M in October 2026, about $18M a year, and that is before removing the wash trades. Five thousandths of one percent of stablecoin payments.

It is still barely there. Whatever this becomes, it has not happened yet. That is the honest starting point for any prediction.

Uber took $193 billion of bookings in 2025 across more than 70 countries, and paid most of it back out to drivers, couriers and restaurants. Here are sixteen of its payments, sorted with the test from the start of this article: who decides, and when.

Leases, insurance, tax and the incentive budget are decided by people and signed by people. Nobody hands a lease negotiation to an agent with a wallet.

Charging the rider's card, paying drivers every week, surge pricing and fraud holds run on rules engineers wrote years ago. More than 75% of bookings are paid by card or digital wallet. That is automation, not agent payments, and it moves the bulk of the money.

A payment changes colour only when a model starts deciding at the moment of paying, inside limits someone set. Uber hedges currency risk with derivative contracts today, on a policy set in advance, with $3.7B of its $7.1B cash held abroad. A treasury agent would read the news and rebalance within limits instead.

So would a courier's agent that cashes out in stablecoins when the rate is right, refunds settled within policy, and an assistant that books and pays for the ride.

Once the buyer is a machine, new payments appear. A rider's agent pays Uber's dispatch agent, a robotaxi buys its own charge, a delivery robot pays a building to open the door.

These only exist if someone outside Uber says yes. A robotaxi charging at an Uber depot is an internal cost, not a payment. It becomes one at another company's charging station, a building's door or a remote driver's console, and that needs thousands of counterparties to accept machine payments first.

Count dollars and the picture tilts back. Almost everything Uber moves today sits in the first two colours. The agent columns are where growth would come from, if the demand shows up.

Visa alone runs about 700 million payments a day. Agents run about 140 thousand.

Every square here is drawn to scale by area.

Consultancies, banks and analyst houses have all published how big agent payments will get. Most of them count something broader than this essay does. So start with the narrowest forecast, and widen from there.

Four Pillars, a crypto research firm, modelled how much agents themselves could pay over x402 by 2028, with input from one of x402's authors at Coinbase. It is the only forecast I found that counts what this essay counts: payments an agent decides and sends.

Agents mostly buy data, search and inference by the call, inside companies. About $186B a year by 2028. Big for crypto, still a rounding error for payments.

Agents restock inventory, book freight and pay suppliers within a budget. About $686B a year, nearly twice what stablecoins carry in real payments today.

Agents decide what to buy to reach a goal, and pay other agents, robots and vehicles for every step. About $1.9 trillion a year by 2028.

Morgan Stanley ($190B to $385B) and Bain ($300B to $500B) forecast US goods bought through agents by 2030. That includes a person approving the checkout, and most of it is paid by card. An agent helped, but it did not necessarily pay.

Juniper ($1.5T), Edgar Dunn ($2.9T) and McKinsey ($3T to $5T) count purchases agents start, influence or orchestrate around the world. These are shopping numbers, and the payment usually still runs through a card a person owns.

Gartner expects agents to sit in the middle of $15T of business buying by 2028. That means an agent took part in the purchase somewhere, not that it paid. It is also the number Four Pillars' model starts from.

The wide lines are the prize. The narrow ones are what actually settles as an agent payment, in stablecoins or anything else. The gap between them is the market the payment rails are still competing for.

Every few decades finance invents a market that did not exist the year before. Grain futures in 1865, the ETF in 1993, credit default swaps a year later. a16z crypto charted them, and the dots keep getting closer.

From the AMM in 2018 to tokenised perpetuals in 2026, the newest markets exist because anyone can deploy one without an exchange's permission.

When the buyer is a program and settlement costs a tenth of a cent, a market can exist for one sensor reading, one second of GPU time or one kilowatt. Those trades were always too small for a human to bother with, so the markets never existed.

Today agents pay for data, search and inference by the call. Next come machines paying machines. After that come markets nobody can list yet.

Agents already pay for data, search, inference and spare GPU time by the call. These listings exist now in x402 catalogs and agent marketplaces.

As models move onto robots and cars, machines start paying each other: a dock by the kilowatt, a door by the access, a remote driver by the minute.

The obvious things autonomous agents will pay for are the ones they run on: memory, electricity and compute. What they buy beyond that, nobody can list yet, the same way nobody in 1865 could have listed the ETF.