x402 Agent Estimates: How TRM’s Two Tests Sort Payers
Two AI agents buy data through x402. TRM’s tests count one and miss the other: can you guess which?
IN 30 SECONDS
TRM’s agentic test is a behavior test: it labels an x402 payer mostly by how its payments look. Its two versions set the ends of the range, 0.6% and 7.5% of screened value, and both can be wrong.
KEY TAKEAWAYS · 3
- Permissive test: varying amounts averaging under a dollar, via a facilitator. Strict test: that for months, plus an agent-registry entry or several sellers.
- The strict test may miss a single-purpose agent, and a script with varied payments can pass: a heuristic, not a verdict.
- Recent x402 headlines don’t say who decided to pay.
On this page
How TRM’s two tests sort payers
IN SHORT TRM labels each payer by behavior: a permissive test first, then a strict one with extra conditions.
WORDS TO KNOW
- Agentic
- A payment that looks, from outside, like an AI agent chose it. A label from TRM’s tests, which doesn’t say who decided.
- Heuristic
- A rule of thumb that guesses from patterns. It can be wrong both ways.
- Agent registry
- A public on-chain list where an address can be declared as an AI agent. ERC-8004 proposes one; it is still a Draft.
STEP BY STEP
Step 1. The question. TRM reads the payment pattern, plus a registry entry for the strict test.
- The question TRM reads the payment pattern, plus a registry entry for the strict test.
- Looks like a script A repeated fixed price reads like a script; without a facilitator, the payer is out of scope. Not counted.
- Permissive test Varying amounts averaging under a dollar, through a facilitator: the permissive test counts it.
- Strict test The same for months, plus a registry entry or more than one seller: the strict test counts it too.
IN PLAIN WORDS
Result: “between 0.6% and 7.5%” of the USD 25.62M screened value, by value, not by transaction count. The strict test gives the low end, the permissive one the high end.
ERC-8004 is the Agent registry A public on-chain list where an address can be declared as an AI agent. ERC-8004 proposes one; it is still a Draft. See the glossary the strict test checks. It’s still a Draft, sign-up is voluntary, and TRM notes most don’t use it. So the range is a Heuristic A rule of thumb that guesses from patterns. It can be wrong both ways. See the glossary estimate, not a count of verified agents.
Coming up next: So what kind of agent slips past both tests?
A worked example: two agents, opposite verdicts
IN SHORT Two agents can pay for data and get opposite verdicts from the same tests.
A WORKED EXAMPLE
Illustrative numbers
PASSA research agent buys reports from three APIs
INPUT
- Payer
- AI agent (research)
- Sellers
- 3 data APIs
- Prices
- 0.05, 0.08, 0.15 USD
- Period
- 4 months
WHAT HAPPENS
- It passes TRM’s screens, through a facilitator.
- Amounts vary and average under a dollar, so the permissive test counts it.
- The pattern lasts months and spans three sellers, so the strict test counts it too.
OUTPUT
- Verdict
- appears agentic (both tests)
MISSEDA single-purpose agent pays one API all day
INPUT
- Payer
- AI agent (single-purpose)
- Seller
- 1 weather API
- Price
- 0.08 USD, every call
WHAT HAPPENS
- Same screens, same facilitator.
- One fixed price repeats: it reads as a script.
OUTPUT
- Verdict
- not counted, though an agent decided
TRM names this blind spot itself: “many agents today could be single-purpose, paying one service repeatedly, which this test would read as a script.” So the strict test may undercount.
ITEM GET
A heuristic, not a verdict
Recent x402 headlines, next to the data
IN SHORT Recent x402 headlines count members and payments, not who pays.
None of this contradicts TRM, because none of it measures the same thing.
BlackRock’s paper argues that broad AI adoption “may represent an underappreciated source of demand” for digital assets. That’s their thesis, quoted. It isn’t a measurement, and I don’t endorse it.
FAQ
It’s the open-governance body, under the Linux Foundation, that now looks after x402. Coinbase contributed the protocol. It launched on 14 July 2026 with 40 members, including Coinbase, Cloudflare, Google, Visa, Mastercard, Stripe, AWS and Circle.
Not on their own. On 2 October 2026, t54 reported that the XRPL (the XRP Ledger blockchain) passed 10 million x402 payments through its facilitator. That count gives no total value and doesn’t show which payments AI agents made.
Sources: TRM Labs, Linux Foundation, Block, BlackRock, t54, ERC-8004 (Draft).
STAGE CLEAR
What you now know
- What each of TRM’s tests looks at, and which end of the range it sets.
- Why the tests can miss an agent and mistake a script for one.
Check yourself
Tap or click a question to see the answer.
A script pays five data APIs varying amounts, under a dollar on average, for six months, via a facilitator. How could it be labelled?
Agentic, by both tests: varying amounts averaging under a dollar, months of activity, several sellers. The tests judge behavior, not who decides.
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.