TRM Labs, the blockchain-intelligence firm, went looking for the AI agent economy inside Coinbase's x402 payment protocol and came back with a smaller number than the narrative promised. Examining $52.7 million across roughly 198.9 million settlements, TRM's analysis attributes only 0.6% to 7.5% of that value to addresses that actually behave like agents — not scripts hammering a single endpoint. The infrastructure for machines paying machines is real and growing. The machines mostly aren't the ones spending yet.
What TRM actually measured
x402 is Coinbase's 2025 standard for folding a payment into an HTTP 402 response, live on Base, Solana, and Polygon since May 2025 — the plumbing meant to let an agent (or anything else) pay per API call without a subscription or a human at a checkout page. TRM pulled the settlement data, and USDC did almost all of the work: $52.47 million of the $52.7 million total, 99.6% of settled value. After stripping anomalous flows — the kind of traffic that looks more like test transactions or wash activity than commerce — $25.62 million remained for TRM to actually classify. Agent-attributable value came out of that pool at 0.6%–7.5%, depending on how generously "agent" gets defined.
The tell: a script hits one door, an agent tries a few
The interesting part isn't the total, it's the method. TRM's logic, in its own words: "a true agent explores across multiple services and products, while an address repeating the same price behaves more like a script hitting one service over and over." That's a genuinely useful heuristic — an agent making purchasing decisions should show some variance in what it buys and from whom; a wallet hitting the same fractional-cent endpoint tens of thousands of times reads more like a cron job than a decision-maker. Most of the wallets TRM looked at fell into that second bucket.
TRM flagged its own blind spot too: a single-purpose agent that only ever pays one service would get misclassified as a script under this same test. So the 0.6%–7.5% range is a floor built on a specific, defensible definition — not a hard ceiling on how much of x402 is genuinely agentic. Worth remembering the next time someone quotes the low end as settled fact.
This doesn't kill the thesis. It prices it correctly.
Amazon, Binance, and Coinbase are all still building toward agents that pay for things without a human clicking "confirm." That work continues regardless of this week's ratio. What TRM's numbers say is narrower and more useful: the payment rails for agentic commerce exist ahead of the agentic commerce. That's normal — infrastructure usually gets built before the traffic that justifies it — but it's a different claim than "AI agents are already moving real money online," which is the version that keeps getting repeated as settled fact.
This is exactly the gap the Trust Rule exists to flag, whether it shows up in a payment-rail transaction count getting read as proof of autonomous demand or a listing claiming "Verified" with nothing behind it. Self-reported and verified are not the same sentence. TRM did the unglamorous work of actually looking at behavior instead of counting wallets. More of that, please.
What to watch
Watch whether TRM's own recommendation — better agent registration and reputation systems built for high-volume, low-value transactions — gets picked up by anyone running an x402 facilitator, since that's the actual fix for the classification problem TRM had to work around by proxy. Watch the agent-share ratio itself over the next few quarters as more products ship agents with real wallets and spending logic, rather than a wallet with a script bolted on. And watch for the next dataset that runs this same test against x402 traffic on Solana and Polygon specifically — this analysis leaned heavily on the USDC/Base-shaped bulk of the volume, and the ratio elsewhere could look different.
Sources
- ▸[AI Agents Spending Money Online? New Research Says Not Really — Decrypt](https://decrypt.co/378103/ai-agents-spending-money-research)