Loading
What agents buy from other agents, settled on chain.Note 23
1,106 addresses; 0.5% of everyone who has ever paid on this rail; carry 79% of its payments, while 63% of payers paid exactly once and never came back. The rail is not a crowd of buyers; it is a few dozen programs and a long tail of people who tried it.
None of this proves an agent, and it is not meant to. The rail publishes authorised transfers; the log carries no statement about who authorised them. A cron job is automated and is not an agent, and a person with a script is both and neither. What can be separated is “behaves like a program” from “we cannot tell”, and nobody is awake for 22 hours of the day across 3 days.
The registration is the only claim here and it is nearly absent. 462 of 221,544 payers hold an ERC-8004 identity, and 4 of the 1,106 that never stop. The addresses that behave most like agents are almost entirely absent from the registry built to list them.
One class here is not behaviour at all, and it is the only one that could be. An attested payer is an address somebody put their name to in a signed declaration we hold and can re-check against the text at any time. It is the strongest attribution this rail can carry and it currently covers 0 of 221,544 payers. The class exists so the figure can be read the day it changes, rather than being built once somebody asks for it.
Where the class boundary falls is a choice, so the distribution it was cut from is below. Payments are counted, not valued: this rail settles in several tokens and summing them would be adding different units.
x402
221,544 addresses, 5,632,731 payments. Counted, never summed: the rail settles in several tokens.
x402
The same rail counted by address rather than by payment. The two rings are near mirror images, which is the finding: the addresses are a crowd and the payments are not.
A buyer counts as returning at 7, 30, 90 days if they paid again at or after that point — not merely twice, since two payments an hour apart say nothing about whether anyone is still there a month later.
| Window | Came back | Could have | Share |
|---|---|---|---|
| 7 days | 16,389 | 133,640 | 12% |
| 30 days | 1,791 | 12,181 | 15% |
| 90 days | 637 | 9,696 | 7% |
The three windows are not one cohort shrinking. Each is a different population — only buyers who arrived early enough can be asked the ninety-day question at all, and there are far fewer of them. So a later window reading higher than an earlier one is not a contradiction; it says the people who arrived first behave differently from the crowd that followed, which is a finding rather than an error to smooth away.
The shape the classes were cut from. Read the last two rows first: 51 addresses account for 72% of everything on the rail.
| Payments made | Payers | Their payments | Share of the rail |
|---|---|---|---|
| 1 | 138,898 | 138,898 | 2% |
| 2-9 | 70,999 | 206,046 | 4% |
| 10-99 | 9,550 | 232,528 | 4% |
| 100-999 | 1,888 | 560,948 | 10% |
| 1k-10k | 158 |
| Class | The test | Payers | Payments |
|---|---|---|---|
| Once | Exactly one payment, ever. Cannot be a running service. | 138,898 | 138,898 |
| Occasional | More than one, but not round the clock. A person could do this. | 81,540 | 1,023,334 |
| Always on | 20+ payments across 22+ hours of the day, over more than 3 days. | 1,106 | 4,470,499 |
| Registered | Holds an ERC-8004 identity. A claim rather than a behaviour, so it cuts across the three above rather than sitting beside them. The registry against every rail. | 462 | 7,672 |
| 421,328 |
7% |
| 10k+ | 51 | 4,072,983 | 72% |
The ten largest payers alone carry 1,515,649 payments, 27% of the rail. Counted 9 Sep 2026, 15:06.