How Agentic Is Agentic Commerce? A Population-Scale Measurement of x402 Adoption and Authenticity
This paper reveals that the hundreds of millions of x402 settlements, often cited as proof of a thriving AI agent economy, are largely manufactured and fictitious, with the vast majority representing internal or sponsor-driven transactions rather than genuine, independent agent-to-agent commerce.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a bustling digital marketplace where AI agents are supposed to be the new customers, buying data and computing power with their own money. The headline news? The register says they've made 136,708,672 purchases, totaling $44,121,383.81. It sounds like a gold rush for AI commerce.
But this paper pulls back the curtain to reveal that the register might be counting the same coin over and over again.
The Magic Trick: Counting Ghosts
Think of the x402 protocol like a magic show where the magician (the "facilitator") pays for the tickets. Because the magician covers the cost of the show (the "gas" fees on the blockchain), a producer can make an AI agent "buy" something millions of times for almost zero cost.
The paper argues that the massive number of transactions isn't proof that AI agents are shopping; it's proof that the system is easy to fake. It's like a carnival game where you get a ticket for every time you throw a ball, but the game operator gives you the balls for free. If you throw a million balls, you get a million tickets, but that doesn't mean a million people showed up to play.
The Three Layers of the Truth
The researchers didn't just look at the total count; they acted like detectives, tracing the money to see who actually received it. They sorted every single transaction into three buckets:
- The "Self-Hug" Bucket (Fictitious): About 21.20% of the transactions were proven to be fake. In these cases, the AI agent paid itself, or the money went in a circle and came right back to the same pocket. It's like paying your own allowance to yourself and counting it as income. This alone accounted for 54.08% of the total money moved!
- The "Family Reunion" Bucket (Internal Settlement): Another 63.78% of the transactions happened inside a closed group of wallets controlled by the same operator. The money moved from one family member's pocket to another's, but never left the house. While this isn't a direct "self-hug," it's still internal accounting, not a sale to a stranger.
- The "Real Customer" Bucket (Unattributed): This is the only part that might be real. It represents 15.02% of the transactions. However, even here, the researchers can't be 100% sure. They can only say that this money reached someone outside the operator's control.
The Real Numbers: A Tiny Slice of Pie
When you strip away the magic tricks and the family reunions, the picture changes drastically.
- The total "hype" number is $44,121,383.81.
- The amount that definitely didn't leave the operator's cluster is $23,862,637.72 (the fake stuff).
- The amount that might be real is bounded between a floor and a ceiling.
- The Floor: Only $187,861.35 demonstrably reached a named, advertised service that we can point to.
- The Ceiling: At most, $20,258,746.09 (which is 45.92% of the total value) is not provably manufactured.
So, the "genuine" economy is somewhere between $187,861.35 and $20,258,746.09. That's a gap of two orders of magnitude. The paper concludes that the real, independent demand is likely much closer to the tiny floor than the huge headline number.
The "Star Forest" vs. The Marketplace
If you look at a real marketplace, you see a web of connections: many buyers, many sellers, and money flowing in all directions. The researchers found that x402 looks nothing like that. Instead, it looks like a "forest of stars."
Imagine a few giant, glowing stars (the operators) in the middle. Around each star, there are thousands of tiny dots (the AI agents) that only talk to that one star. They don't talk to each other. The money flows from the tiny dots to the star, and then almost immediately flows back out to the same star's other pockets. It's a closed loop, not a market.
Why Did They Do This?
The paper suggests this isn't an accident; it's a game of incentives.
- The Reward: There are leaderboards and tokens that reward people for having high transaction counts.
- The Subsidy: The "facilitators" pay for the gas fees, making it nearly free to generate these numbers.
- The Result: Operators are essentially "manufacturing" activity to climb the leaderboard. The paper calls the transaction count a "Goodhart metric." This is a fancy way of saying: "When a measure becomes a target, it ceases to be a good measure." The operators are hitting the target (high counts) so hard that the target no longer tells us anything about real adoption.
The Bottom Line
The paper doesn't say AI agents can't buy things. It just says that the current numbers—136 million transactions—are mostly a mirage created by operators recycling their own money to look busy. The real, independent economy is small, bounded, and hiding in the noise. The headline count measures the incentive to transact, not the adoption of the technology.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.