x402 Payments: How Much Value May Come From AI Agents?
An AI agent and a timer script leave the same x402 payment record. So how can anyone estimate how much value comes from agents?

IN 30 SECONDS
x402 is an open protocol, built with AI agents in mind, that lets a website or API charge inside a web request. TRM Labs estimates that 0.6% to 7.5% of screened x402 value looks like agent activity. It’s only an estimate, because a payment record can’t show who decided to pay.
KEY TAKEAWAYS · 3
- TRM Labs counted roughly USD 52.7M across 198.9M x402 settlements through known facilitators since May 2025. Its screens left USD 25.62M of commerce.
- TRM’s tests classify 0.6% to 7.5% of that screened value as appearing agentic: a heuristic estimate of value, not a count of verified agents, not a share of the USD 52.7M.
- The decision to pay happens on the buyer’s side, off the chain, so the payment record alone can’t separate an agent from a script.
On this page
The problem: a payment can’t say who decided
IN SHORT A payment record shows that money moved, not who or what decided to move it.
WORDS TO KNOW
- x402
- An open protocol that lets a website or API (an interface programs use to reach a service) charge inside a web request.
- AI agent
- A program that uses an AI model to choose its next action, such as paying the asked price or not, without a person approving each step.
- Script
- A program that follows fixed rules written in advance, such as paying the same price every hour.
- Facilitator
- A service that checks a buyer’s signed payment and submits it to the blockchain for the seller, often paying the network fee.
- Screened value
- The payment value TRM Labs kept after removing noise such as self-payments, bulk flows and sellers with under 10 buyers.
- Agentic
- A payment that looks, from its behavior, like an AI agent chose to make it. A label from TRM’s tests, which doesn’t say who decided.
x402 An open protocol that lets a website or API (an interface programs use to reach a service) charge inside a web request. See the glossary is often pitched as the way AI agents will pay online. So how much x402 money comes from AI agent A program that uses an AI model to choose its next action, such as paying the asked price or not, without a person approving each step. See the glossary? TRM Labs, a blockchain analytics firm, tried to measure it in a study published on 9 September 2026.
ADAM says: A line on your card statement looks the same whether you clicked “buy” or a subscription renewed itself. But your bank can ask you which was which; a public blockchain has no one to ask.
TRM’s answer: only a small share looks Agentic A payment that looks, from its behavior, like an AI agent chose to make it. A label from TRM’s tests, which doesn’t say who decided. See the glossary, and the range around it is wide. For the x402 basics, see episode 3 on agent payments.
Where I stand: I co-author ERC-8203 (second of three), a Draft in open PR #1614, not merged, covered in stage 3.
How one x402 payment works, step by step
IN SHORT The server names a price, the buyer decides on its own machine, and only the payment lands on the blockchain.
STEP BY STEP
Step 1. Request. The buyer, an agent or a script, asks a paid API for a resource.
- Request The buyer, an agent or a script, asks a paid API for a resource.
- Price The API answers with the code 402 Payment Required and a price.
- Decision The buyer accepts the price, by a fixed rule or by an agent’s reasoning, on its own computer.
- Signed retry The buyer sends the request again with a signed payment authorization.
- Check The API asks the facilitator to check the authorization.
- Settlement The facilitator puts the payment on the chain and pays the network fee.
A Facilitator A service that checks a buyer’s signed payment and submits it to the blockchain for the seller, often paying the network fee. See the glossary does the blockchain work for the buyer. The settlement, the confirmed payment, is then public. TRM sums up what that record can’t show:
“Agency cannot be read off a transaction, because a scheduled script and an agent produce the same record.”
x402 was designed with agents in mind, but nothing in it requires one. A Script A program that follows fixed rules written in advance, such as paying the same price every hour. See the glossary on a timer can walk the same six steps. Only step 3 differs, and it stays off the chain.
Coming up next: So if the chain can’t tell, how did TRM get a number at all?
The number: 0.6% to 7.5% of the screened value
IN SHORT TRM’s range is a share of the value left after screening, not of all x402 money, and not a count of agents.
TRM started with everything settled through known x402 facilitators on the Base, Solana and Polygon blockchains since May 2025: “roughly USD 52.7 million across 198.9 million settlement transactions”.
It then removed self-payments, bulk flows from one or two payers, and sellers with fewer than 10 distinct buyers. About half the value fell away. What was left is the Screened value The payment value TRM Labs kept after removing noise such as self-payments, bulk flows and sellers with under 10 buyers. See the glossary: USD 25.62M.
IN PLAIN WORDS
It’s a share of value, not of transaction count, and a heuristic (a rule of thumb), not a count of verified agents. It can err both ways: the strict test may miss single-purpose agents, and a varied script can pass if it meets the other criteria (stage 2).
ITEM GET
Quote it on USD 25.62M
TRM gives one more, separate figure: the agents it observed on x402 moved roughly USD 5,000 to 11,000 a month through 2026. That’s a monthly volume, not a share of the screened value.
Sources: TRM Labs (9 September 2026).
STAGE CLEAR
What you now know
- How an x402 payment runs: price, decision, signed retry, settlement, with the decision kept off the chain.
- Which total TRM’s range applies to, and what it doesn’t count.
- Why a settlement record can’t tell you whether a script or an agent decided.
Check yourself
Tap or click a question to see the answer.
A headline says AI agents paid 7.5% of the USD 52.7M that went through x402. What’s wrong with it?
Two things. The range is a share of the USD 25.62M screened value, not of the USD 52.7M total. And 7.5% is the permissive test’s high end: a heuristic estimate, not a count of verified agents.
Former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft. Building production AI and regulated Web3 systems.
Have a project to ship or a role to fill?
Discuss your project or roleFollow the engineering notes
Get new articles on production AI and smart contracts by email.

