Agents manage money, buy work from each other, and register identities that claim things. This reads all three from their own sources and publishes what can be checked, including where the answer is that we cannot tell you. That is most of the reason the numbers here differ from everybody else’s.
$10.76B held across the 2,526 of 3,337 open vaults we can put a size on. In 300 of them, holding $4.81B, we saw automated execution and nothing is declared; 1,846 holding $1.01M are declared or sit on a venue built to run software. Automated execution is a machine at work, not proof that a model decides. What the capital is invested in, what depositors actually earned, and how much of it cannot be withdrawn today.
Capital
Every open vault we read, agent-managed or not. A chain absent here is one we do not observe, not one with nothing on it.
405,787 agent-to-agent jobs read, 8,634,443 payments from 313,540 addresses. What agents buy from each other, whether it settles, and what happens to the money afterwards.
Agent work
Jobs bought, by the service named in each opening request. Concentration here is a fact about what the rail is used for, not about who uses it.
Delivery
405,787 jobs read. 150,788 have a phase we do not read, which says nothing about whether they progressed. Among the 254,999 whose outcome we read, 47% completed.
552,680 identities registered under ERC-8004, and what any of it is worth. Registering is a claim; being rated, being reachable and being who you say you are on another chain are checks, and they do not agree.
Registrations
Registering is a claim, not a record. An identity registered on a chain where we read no market has nothing we can link it to, which is why where anyone was rated is the ring that matters and it has a different shape entirely.
Ratings
Ratings are not a crowd: the best-covered chain here is a handful of accounts, and the table on the page below carries how concentrated each one is.
| When | Event | What happened | Size |
|---|---|---|---|
| 2h | A seller repriced | 30 agent services changed their published price. | |
| 2h | Jobs quoted |
Every figure here is read from a chain, a venue’s own indexer, or a registry, and carries the window it covers and the share of the question it answers. Where a number cannot be established it says so instead of printing a zero. An absence and a refusal are different facts, and most of what is published about agents collapses them.
How everything here is measured·What we cover and what nobody can·Data, API and pricing
Net investor return
3,686 depositors across 233 vaults. The capital as a whole earned 9.5% over their own holding periods. An advertised rate cannot contain the split.
Trading
Every open position on GMX, both chains, 3,830 of them, counted rather than sampled. 144 are levered 25 times or more.
Stuck money we can trace
Capital supplied into lending markets that have nothing free, across every lending vault we read, including vaults that can still leave through their other markets. That is a different population from the capital in vaults that cannot be exited as a whole, and neither is a subset of the other. Morpho and Lista markets carry one collateral each, which is why these are exact; Euler vaults accept many and are here only where we traced the pledges.
x402
313,540 addresses, 8,634,443 payments. Counted, never summed: the rail settles in several tokens. This is payments by class of payer, not payers.
Afterwards
Each bar is a share of the same 132,122 addresses paid on a rail. They are not exclusive and are deliberately not drawn as parts of one circle.
Nothing we can see: 106,291 · 80%, the only figure here that is a partition of the rest. It is a fact about how far this linkage reaches, not a fifth thing agents do: revenue moved to an exchange, to an unlinked treasury or onto a chain we do not read looks identical from here to revenue that never moved.
ERC-8004
223 claims name a different chain and this ring is only those; 15 name a registry we do not read, so the rate in the centre leaves them out. The other 4,909 name the registry the card is already in.
| 111 new jobs on the ACP rail, none of which paid anything. |
| 2h | Ratings written | 572 ratings written on base. 29,835 of 95,150 registered agents there have ever been rated. |
| 2h | Money out | Steakhouse Prime USDC: paid out $1.14M, 17% of what it held | $1.14M |
| 2h | Money out | Felix USDT0: paid out $0.74M, 17% of what it held | $744.0k |
| 2h | Money out | Yearn OG WETH: paid out $3.63M, 42% of what it held | $3.63M |
| 2h | Money in | August USDC V2: took in $1.16M, 11% of what it held | $1.16M |
| 2h | Money in | WETH ARM Vault: took in $1.60M, 19% of what it held | $1.60M |