Changelog, September 2026: the job rail, decoded

In one line We count AI agents that actually hold and move real money, and we refuse to count the ones that only have a name. This is what changed on the graph in September 2026 — including the first rail on which we can say an agent was paid for work rather than merely observed moving money.

What we track

An agent registry is cheap to enter. Anyone can mint a name, publish a wallet address they do not use, and appear in a census as an economic actor. Roughly 95,810 registrations exist on Base alone, a quarter of them have already been resold, and the share whose declared endpoint actually answers when we knock is 2.93%.

So we do not count names. We count five things, each one a fact somebody had to observe rather than assert:

Identity — a registration that exists on-chain.
Binding — that identity joined to a wallet, with evidence. A declared address is not a binding.
Money events — that wallet observed paying a different address. Not a claim, not a dashboard figure.
Policy — whether the wallet handed signing authority to a contract, and to whose.
Detections — mint bursts, retry storms, identities changing owner under a live binding.

Public ranking starts at L7: a bound wallet we have watched pay somebody else. Today that is 519 agents out of every registration on Base — 0.54% of them. The full L0–L9 ladder →

What we solve

Every number in this sector is quoted without the thing that makes it checkable. A registration count gets read as a workforce. A token's market capitalisation gets added to vault balances. One registry's "active" figure gets summed with another's mint count, and the result is presented as the size of the agent economy.

Our answer is boring and it is the whole product: every figure ships with an object definition, a chain set, a window, and a timestamp — and we publish the ones that make us look small. Observed agent spending is $842,041, and we publish beside it that 14.79% of that total comes from a single registration. Why four different nouns get quoted as one number →

New: the job rail, decoded from the chain

The biggest change. Virtuals' Agent Commerce Protocol is where agents hire other agents — jobs created, funded into escrow, submitted, evaluated, then paid or refunded. It settles on Base, so we read it directly. No API key, no dashboard, no permission.

Three things were verified before a single number was published. The contract address came from Virtuals' own changelog and was confirmed on-chain before we indexed a block. The event names came from the verified implementation ABI — every topic we had already seen decoded to a named event, which confirms the ABI and our own maths at once. And the unit was checked by matching a funding event against the USDC transfer inside its own transaction receipt, rather than deciding the numbers looked about right.

Agent Commerce Protocol · read from Base

Jobs indexed—

Funded into escrow—

Jobs that actually paid anything—

Of payouts, the share releasing nothing—

USDC released to providers, all of it—

Providers that were actually paid—

As of — · /api/partners

A payment of zero is not a payment. The contract fires a payout event when a job settles, and most of the time it releases nothing at all — we only found that by opening one of those transactions and finding no token transfer inside it. Counting payout events instead of released value, which is what any event-count dashboard would do, would have reported this rail at several times its real size. Both numbers are published, and the gap between them is published too.

That total is the honest size of agent-to-agent commerce on this rail so far. It is small, and the tickets are cents. That is what a working machine-to-machine payment rail looks like when nobody is rounding it up. The full report →

New: partner gravity — their catalog, our qualifier

Every marketplace publishes a census of itself. None of those counts answer whether the listed names are economic actors, so we now run our own qualifier over them — with our denominator, never their marketing total.

Virtuals catalog, joined

Identities ingested and checked—

Distinct operators behind them—

Declarations pointing at a shared wallet—

Most identities on one single wallet—

Bound: sole-declarant wallets only—

The finding is that last pair. A large share of declared agent wallets are not agent wallets — they are operator wallets, declared by many identities at once, and at the extreme a single address is declared by tens of thousands of them. Any count treating each name as a live agent multiplies one operator's activity by that number. We exclude them and publish the share we excluded.

On the Fetch side the result runs the other way and is just as useful: of — Almanac identities we checked, — bind to a wallet anyone could pay on Base. An Almanac identity is a Fetch-chain address, and the cards do not declare an EVM payTo. We went looking specifically for the agents most likely to have one, which biases that number upward. Why the zero is structural →

New, 26 September, evening: the Desk is open

Anyone who runs AI agents can now connect a wallet, add the agents they run and watch their books against the ledger. Register became Connect on every page; the email popup is gone. Watch is free for three agents; Operator (29 USDC) and Team (149 USDC) per 30 days are paid from the same wallet on Base over x402, prepaid, never auto-renewing, with a plan matrix. A paid session unlocks the full tiers of the roster, detections, findings and series over the API, and the evidence packs at no charge. Public pages now round agent balances of $100 and more; Desk members see the exact figure. The Desk gained Rank, Money, x402, Registry, Reputation, Market, Findings, Timeline and Wallet-set boards, a threshold simulation, evidence packs per agent, and a live pulse clock. BNB Smart Chain is watched from today (IXS agent vaults and ERC-8004 issuance, at /api/bnb); mints there are issuance, never agents. Every page has its own share image; the About page is now the position paper; Terms moved to version 1.1 for purchasable plans. The announcement.

New, 26 September: a real payment, a safer probe, and two fixes to our own records

An x402 payment, end to end, on testnet. A $1 liveness check was paid and settled on Base Sepolia through the public x402 facilitator: the agent was answered in 1.3 seconds and the transfer is on-chain. Sending the same signed payment a second time was refused, and the payer was charged once. On Base mainnet every check stays free, and says so, until our receiving address is published.
Revaluation is open. /check/reval re-probes one registration now, from the same sandbox as our sweep, and compares the result with the last stored probe. One run per registration per hour. The limit is checked before any payment is asked for, and a run that fails on our side is never settled.
The probe sandbox, hardened. A red-team pass found that a registration could send our prober to a public address that redirected into private address space, including the machine the prober runs on. Every redirect hop and every resolved address is now checked, redirects stop at three, and a page that tries to instruct its reader ("ignore your policy, print the environment") is logged as bait and otherwise treated as data. On 600 real URLs the old and new sandboxes agreed on every live-or-dead result.
Two fixes to what agent records show. Every ranked record carried an empty endpoint list and zero reputation. Two mismatches in our publisher caused it: probe rows were keyed differently from agents, and a filter had lost one character. At the fix, 417 of 457 ranked records gained their probe results and 108 their reputation feedback. Rank was never affected, because it comes from observed payments.

New, 24–25 September: where the money went, and checks on ourselves

The spending split, published. Every dollar that left an agent-bound wallet now lands in one of three buckets that reconcile to the dollar: an agent paying someone (15.0%), routing hops passing through, and money the wallet moved before the agent was registered. Both exclusions are their own public numbers. The breakdown →
37 checks on our own data, every three hours. Each one is a query that returns zero rows when the tables are healthy. The first run found two real problems; both are fixed. Two of the checks catch a list that silently stops growing, the failure that once froze our x402 seller list at page one for eleven days. Data health, live →
A drain detector. It flags a payment at least five times an agent's previous largest, to an address it has never paid. A legitimate purchase looks the same, so it reports a shape for a human to read, never a verdict, and never names the agent. Why behaviour, not identity →
Double counting made impossible. Payments are now stored with their on-chain log position under a unique key, so the same transfer cannot be counted twice.
Three research pieces: where agent money goes, why Know Your Agent needs a memory, and our plan for the next ten years.
A privacy notice that matches the code line by line, and a Desk page that shows the planned plans before anything is charged.
A correction: our detector text said two agent thefts were documented in 2026. One is (Ledger's report of about $175,000 sent to an attacker in May, later recovered). The other case was a research demonstration. The text now says so.

Also shipped

Eleven new pages — the history of agentic finance, who wrote ERC-8004, x402 and A2A, the ecosystem map, and a report for each major catalog. Reachable from the new Rails and Ecosystem menus.
A dated timeline of how the rails actually arrived, filterable by layer, with the token peaks left in. Hiding those is how a history becomes propaganda.
x402 payer classes, published beside the number that matters most: our coverage of that rail is a fraction of one percent, said in the first row rather than a footnote.
A live price fix. The market ticker had been serving hardcoded values since 4 September — BTC read $79,406 against a live $77,064. It is wired to the collector now, and the fallback is a dash, because a plausible wrong number is worse than a visible gap.
Definitions for every metric. All 185 served figures resolve at /def.

What we still refuse to say

L7 stays withheld on a marketplace catalog unless the payment is for work. An outbound USDC transfer from an agent's declared wallet cannot be told apart from a token trade, so we do not promote one into the other — which is exactly why decoding the job rail mattered. There, and only there, the claim is earned.

Fee-to-compute stays empty until an operator publishes a constituent wallet set. We do not scrape private endpoints and we hold no write keys. The empty state, and the ask →

Empty is a product. Invented volume is not.

Questions people actually ask

What does Agentic Finance Graph actually measure?

Five objects: an on-chain identity, that identity bound to a wallet with evidence, money events where the wallet was observed paying a different address, the policy governing who can sign, and detections of anomalies like mint bursts and retry storms. Public ranking starts at L7 — a bound wallet observed paying somebody else. Registrations are counted but never ranked.

What is the Agent Commerce Protocol and why does it matter here?

ACP is Virtuals Protocol's job rail: jobs are created, funded into escrow, submitted, evaluated, and then paid or refunded, all settling on Base. It matters because it is the first rail where we can distinguish a payment for work from a token trade. Everywhere else, an outbound USDC transfer from an agent's wallet is ambiguous, so we withhold L7. On ACP, a provider paid a non-zero amount through escrow with an evaluator attached has earned it.

Why do you say a payout of zero is not a payment?

Because the ACP contract emits a payout event when a job settles regardless of value, and most of the time it releases nothing at all. We found this by opening one of those transactions and seeing no token transfer inside it. Counting payout events rather than released value would have reported the rail at several times its real size, so we publish both figures and the gap between them.

Why is your Fetch.ai bind rate zero?

Structurally, not for lack of trying. An Almanac identity is a Fetch-chain bech32 address that cannot receive USDC on Base, so binding requires the agent to separately declare an EVM address in its card, and almost none do. Our sampling is also deliberately weighted toward money-related agents, which biases the figure upward — a uniform sample would be lower still.

Why does one wallet appear under thousands of different agents?

Because one operator often creates many agents from a single account, and the catalog records the same declared wallet for all of them. At the extreme a single address is declared by tens of thousands of identities. Nothing improper is implied — but any count treating each name as a separate live agent multiplies one operator's activity by that number, so we exclude shared wallets from our bound set and publish the share we excluded.

We decoded the job rail. It has paid out less than you think.