BTC —◆ ETH —◆ SOL —◆ STABLECOIN FLOAT $313.17B◆ NATIVE USDC ON BASE $4.31B◆ DEFI TVL $95.53B◆ THE PUBLIC LEDGER OF MACHINE MONEY — AGENTICFINANCEGRAPH.COM◆

Home/Research/Where agent money goes

Research · measured on Base

Only 15.6% of AI-agent money is real spending

We followed every dollar that left a wallet bound to a registered AI agent on Base. Most of it was not an agent buying anything.

Agentic Finance Graph · Published 2026-09-25 · Measured 24 Sep 2026 · Base mainnet · ~1,400 words

Direct answer Of $4,856,774 that left wallets bound to registered AI agents on Base, $757,082, or 15.6%, was an agent paying someone. 31.6% was routing hops passing through agent wallets, and 52.8% was money those wallets moved before the agent was even registered. Current figure: 15.0% of $5.62M.

This week the phrase "machine-native economy" is everywhere. BlackRock put its name on it, and every feed has redrawn the same diagram: a human paying a merchant today, an agent paying an API tomorrow. The rails behind that picture are real. x402 lets a web service charge per request. ERC-8004 gives agents an on-chain identity. Stablecoins settle in seconds.

What almost nobody publishes is a careful count of what agents actually do with money. So we built one. For every AI agent registered on Base, we follow the wallets bound to it and record every dollar that leaves them. Here is what that count shows.

$4.86 million out, $757,000 spent

Since the first agent registrations on Base in early February, $4,856,774 in USDC has left wallets bound to registered agents. It splits into three parts that add up to the dollar:

Every dollar that left an agent-bound wallet on Base, measured 24 Sep 2026.
What the money wasAmountShare
An agent paying someone else$757,08215.6%
Routing hops: money passing through an agent wallet on its way somewhere else$1,533,61131.6%
Money the wallet moved before the agent was registered$2,566,08152.8%

Only the first line is an agent spending money. The second is plumbing: the same dollars bounce through a wallet as part of a swap or a relay, and counting each hop as a purchase counts the same money several times. The third is history. Many agents are bound to wallets that existed long before the agent did, so counting that wallet's whole past as "agent activity" credits the agent with money it never touched.

A total that doesn't separate these three is more than six times too big. That's not a guess about other people's numbers; it is the gap in our own data before and after the exclusions. We publish both exclusions as their own figures, so anyone can check what we left out and why.

95,520 names, 443 payers

Registrations are the other number people like to quote. 95,520 AI agents are registered on Base. 443 of them have ever been seen paying anyone, which is 1 in 216.

A registration is a mint, not an agent. It costs almost nothing, and it can be done in bulk. 28.6% of all registrations were minted in bursts by single addresses, and the ten most active minters account for 35.1% of the whole registry. None of that is wrong, since platforms legitimately issue identities for their users. But a registry count measures how many names were issued, not how many agents are working.

We rank agents by evidence instead of by claims. An agent reaches level 7 (L7) on our ladder when a wallet bound to it is seen paying a different address. L8 adds a funded wallet. L9 needs repeat counterparties. Most of the registry never gets past the name. How the ladder works →

The paying agents are concentrated too

Even among the 443 agents that do pay, the picture is narrower than the headline suggests.

One escrow contract, TermixEscrow, receives 25.5% of all counted agent spending: $193,091 from 214 agents. It's a work-escrow platform, where agents pay into orders that settle when the work is delivered. More striking: 189 of the 443 paying agents made their very first payment to that contract, inside eleven days in August.

That's real economic activity, but it is one platform onboarding many agents at once, not a crowd of independent buyers. When you remove that wave, new paying agents arrive at a steady six or seven a week. The agent economy on Base doesn't grow smoothly. It grows in platform-sized steps.

Reputation shows the same shape. Across 455,286 on-chain reputation events, ten raters wrote 51.7% of all the feedback. A signal that ten addresses can dominate isn't much of a signal.

Why the outside estimates disagree

We aren't the only ones asking whether agent payments are real. TRM Labs reportedly screened roughly half of x402 settlement volume out as self-payments and anomalies. Artemis called the x402 boom "still mostly a mirage" back in February. News outlets report tens of millions of dollars a month.

These numbers are hard to compare because they describe different populations under different rules. x402's public stats show around 10,000 unique payers a day on Base, but among agents that are actually registered on-chain, x402 is only 3.1% of payments by count, and $23.40 by value. Most of the x402 traffic isn't coming from agents with an on-chain identity at all.

Nobody is lying here. The disagreement is a measurement problem: what counts, who counts, and whether the exclusions are published. That's the part we think matters most. Four nouns people add together →

What would make the number honest

Every figure we publish carries four things: what exactly is counted, on which chain, over what window, and as of when. Every metric has a definition id. And before anyone trusts the tables underneath, 37 automated checks run against them every three hours. Each check is a query that returns zero rows when the data is healthy. Every check, live →

We built those checks because every serious correction we have made so far was a data-shape error someone spotted by eye: routing counted as spending, a wallet's history before its agent existed counted as the agent's (that one inflated the headline by 46%), and a seller list that sat frozen at page one for eleven days while every run reported success. Each is now a named check, so the next one gets caught by a machine instead of a reader.

The honest summary

The rails for machine money exist, and they're spreading fast across chains and custodians. The money moving on them, measured carefully, is small: three quarters of a million dollars of real agent spending on Base, from a few hundred agents, a large share of it through one platform.

That isn't a discouraging number. It's the size of the thing that hasn't happened yet, and it's far more useful than a headline total built by adding things that can't be added. If agents do become a real payment channel, the first people to notice will be the ones counting carefully.

All figures: Agentic Finance Graph, Base mainnet, measured 24 September 2026. Live data and definitions: /api/state. Outside figures as reported by Decrypt (TRM Labs), Artemis on X and x402scan.

Questions people actually ask

How much do AI agents actually spend on-chain?
On Base, $757,082 of the $4,856,774 that left wallets bound to registered AI agents was an agent paying someone else: 15.6%, measured 24 Sep 2026. The rest was routing hops and money the wallets moved before the agent was registered. The current figure is at /api/state.
How many registered AI agents actually pay for anything?
443 of 95,520 agents registered on Base under ERC-8004 have been seen paying a different address: about 1 in 216. A registration is a mint, not proof of activity.
Why do estimates of x402 and agent payment volume disagree?
They count different populations under different rules. Public x402 stats include many payers with no on-chain agent identity; among registered agents, x402 was 3.1% of payments by count and $23.40 by value on 24 Sep 2026. Estimates also differ on whether self-payments and routing are excluded.