AI agents already built a market worth nearly $3 billion. But only $1.6 million to $3 million of March activity in the most visible public payment layer appears economically real.
That gap says more about the state of the machine economy than any token chart or product launch.
The public onchain AI-agent economy does not operate as one clean category. It spans agent identity systems, machine-payment rails, and token-driven trading activity. These layers overlap, but they measure different parts of the market. Registered agents, active wallets, payment participants, and token traders do not describe the same thing.
The market already has real infrastructure, but the data still shows a system where real usage competes with noise and incentives. We are only looking at the public, measurable onchain slice of that market, yet even this limited view already shows the core parts of an economic system: identities, counterparties, payment rails, wallets, and transaction flows.
The real question is not whether the AI-agent economy exists, but what kind of economy is forming underneath the narrative.
89,000+ Registered Agents Is Only One Layer
The AI-agent economy is measurable, but only if you stop treating it as one metric. The public onchain AI-agent economy splits into four layers, and each one measures a different part of the system.
| Layer | What it measures | Public proxy |
|—-|—-|—-|
| Identity | Agent registration | ~89,451 registered ERC-8004 agents |
| Payments | Machine-payment participation | ~94,000 buyers and ~22,000 sellers on x402 |
| Usage | Agent-related onchain activity | ~99,000 unique addresses on Base |
| Trading | Speculative/token activity | AI-agent token and DEX volume |
The problem is not that the market lacks data. The problem is that the data describes different layers of the same system.
One source showed about 79,834 agents across 20 active chains, while another March snapshot showed about 89,451 registered ERC-8004 agents and nearly 130,000 cumulative agents since the start of the year.
That is why the market works better as a set of connected metrics than as one headline number. Base shows the same issue.
One public snapshot showed about 21,000 registered agents, 462,000 agentic transactions year to date, and 99,000 unique addresses.
These figures help map activity, but they do not answer the same question. Right now, no public dataset gives one canonical count of “AI-agent wallets”.
From $8.4M to $24M in Three Months
Once we define the market correctly, the next point becomes clear: this system is already growing. The public onchain slice of the AI-agent economy still looks early, but it no longer looks static.
In x402, rolling 30-day payment volume rose from $8.4 million in January to $15.9 million in February and $24 million in March.
Over the same period, transactions increased from 1.18 million to 2.03 million and then to 3.48 million, buyers from 44,800 to 71,200 and then to 109,000, and sellers from 3,180 to 5,420 and then to 8,700.
The identity layer also grew, though more slowly, with registered-agent counts rising from roughly 79,000 in January to nearly 90,000 by late March.
Base data points in the same direction. Base appears as one of the most visible execution layers in public datasets, though that may reflect data availability rather than full ecosystem dominance.
Weekly unique addresses increased from about 92,000 in January to 141,000 in February and 205,000 in March. Weekly agentic transactions rose from 780,000 to 1.42 million and then to 2.31 million over the same period.
In percentage terms, x402 30-day transactions grew about 194.9% from January to March, while Base weekly agentic transactions grew about 196.2%.
These numbers do not describe a mature market, but they do describe a live one. Still, growth alone does not tell us how much of that activity carries real economic weight.
A $3B Market With Only $1.6M–$3M of Real Activity
When only $1.6 million to $3 million of $24 million survives filtering, the machine economy may be real, but its clearest measurable payment core is still tiny.
At the token-market level, the AI-agent sector already looks large. CoinGecko puts the category at about $2.93 billion. But in the most visible public machine-payment slice, x402 showed about $24 million in raw 30-day payment volume in March, while filtered economic activity fell to just $3 million under one approach and $1.6 million under a stricter one.
What Filtering Changes
Filtering changes how we read the market. Raw activity can include circular payments, subsidized flows, self-paid interactions, spam-like transactions, and other low-value behavior that inflates visible usage.
Once those flows drop out, the measurable economic core looks much smaller. In practical terms, about 12.5% of March raw x402 volume survived one filter, while only 6.7% survived a stricter one. That means roughly 87.5% to 93.3% of visible machine-payment volume may not reflect meaningful economic exchange.
Public data still does not show whether the share of economically meaningful activity is improving over time, as a clear raw-to-filtered comparison is only available for March.
A multibillion-dollar token category can coexist with a much smaller base of real onchain commerce.
Real Usage Exists But It Still Doesn’t Prove Real Demand
There is evidence of real reuse, but not enough evidence yet of broad, high-quality economic demand. The clearest signal comes from how intensively participants use the system over time, not just from how many of them appear.
On x402, transactions per buyer rose from 26.3 in January to 28.5 in February and 31.9 in March. On Base, weekly transactions per active address increased from 8.48 to 10.07 and then to 11.27. These patterns suggest repeat usage, not just isolated experiments. In both x402 and Base, transaction growth outpaced participant growth, suggesting that users are returning and doing more than a single transaction.
At the same time, repeat behavior still does not prove economic quality. A system can show rising activity per buyer or per wallet and still depend on subsidized flows, circular activity, or low-value automation. Public data still does not provide clean visibility into retention, cohort behavior, or consistent economic-quality trends over time.
$1.12B in Trading, $24M in Payments – Speculation Still Dominates Utility
The market has already priced in more than the clearest public payment layer has economically delivered.
In one public snapshot, AI-agent-related DEX volume reached about $1.12 billion, while raw machine-payment activity in the clearest public payment slice sat at about $24 million.
In the public comparison, trading activity ran at roughly 47 times machine-payment volume.
These numbers do not show an empty sector. They show a market where trading, token pricing, and attention still run far ahead of productive autonomous commerce.
AI Agents Have Lost $7.55M — But the Bigger Problem May Be Weak Economic Signal
Users are already losing money, but the deeper problem may be how much visible activity still fails to translate into meaningful economic use.
Publicly attributable losses in this slice of the market already total about $7.55 million across agent-related and automated-wallet incidents. That confirms real financial risk, even at this early stage.
But direct losses may still be smaller than the market’s broader signal-quality problem. In the clearest public machine-payment example, March raw x402 volume reached about $24 million, while filtered estimates reduced economically meaningful activity to $1.6 million to $3.0 million.
That leaves roughly $21.0 million to $22.4 million of visible activity in a gray zone between noise, incentives, and economically thin usage.
One public Base snapshot also showed a 14.2% failed or reverted interaction share in agent-related flows. The larger issue is not just hacks. It is that weak economic signal that makes the market look bigger, clearer, and more mature than it really is.
The Most Likely Future Is a Hybrid Machine Economy
The idea that crypto would replace the financial system was always the wrong frame. It did build something transformative, but not what most people expected. Instead of becoming an alternative system for humans, crypto is turning into the financial infrastructure for machines.
Today, this still looks small. Stablecoins and crypto combined account for roughly 0.02% of global payments, around $400 billion annually. At a global scale, this is not even a rounding error. But treating this growth as linear misses the point. Technologies like crypto and AI do not scale linearly, and the shift from marginal to dominant can happen much faster than traditional models suggest.
Over the next five to ten years, the machine economy is unlikely to become a fully crypto-native parallel system, but it will not be absorbed by traditional finance either. The more realistic outcome is a hybrid model. Crypto will remain the execution layer where autonomous systems operate, because it already solves what machines need: programmable payments, contract-based logic, and value transfer without friction.
At the same time, as soon as real capital and real scale enter the system, regulation becomes unavoidable. Large flows, stablecoins, and core infrastructure will operate within regulated environments, while identity, verification, and trust frameworks will become a required part of how machines transact.
This will not look like a fully decentralized cyberpunk economy. It will look like a layered system. Open crypto rails will continue to power experimentation and machine-to-machine coordination, while the parts that reach scale will increasingly integrate with regulated financial infrastructure. What emerges is not a replacement of the existing system, but its extension into a new type of economic activity.
Financial Layer for Machines Is Forming Onchain
The point is not that the AI-agent economy is fake. It is that the part of it we can already measure looks more like an early coordination and payment layer than a fully formed autonomous economy.