Blockchain intelligence firm TRM Labs has looked at what is actually moving across Coinbase’s x402 protocol and reached an uncomfortable conclusion: most of the payment volume is not coming from AI agents. The report, covered by Yahoo Tech, lands in the middle of a funding and marketing cycle built almost entirely on the premise that autonomous agents are already spending money online at scale. The data suggests they are not — or at least not in volumes that justify the word “economy.”
What TRM Labs Found
TRM Labs analyzed on-chain activity associated with x402, the HTTP-native payment protocol Coinbase open-sourced to let software — including AI agents — pay for API access with stablecoins. Their finding, as reported by Yahoo Tech, is that the majority of payment volume on the protocol cannot be attributed to autonomous AI agents. That is a meaningful distinction. A protocol can show rising transaction counts while the actual payers are humans testing integrations, developers paying for inference, or participants in incentive programs. Volume is not the same as autonomy.
The research does not claim agents never transact on x402. It claims the dominant share of measured volume comes from elsewhere. For a sector where pitch decks routinely cite x402 as proof that machine commerce is live, that is the number that matters.
Why the Gap Between Narrative and Data
There are structural reasons agent-initiated payments remain thin. Most deployed “agents” today are assistive workflows with a human approving outcomes, not autonomous economic actors holding wallets. Genuine agent-to-agent commerce requires machine-readable pricing, persistent agent wallets, and a reason for one agent to buy from another — a combination that is still rare outside demos and hackathons. We noted earlier this week that current machine-to-machine flows skew heavily toward sub-cent transactions that barely register in dollar terms. TRM’s data is consistent with that picture.
There is also a measurement problem worth flagging. On-chain analysis can identify addresses, amounts and timing, but attributing a transaction to an “AI agent” versus a script a human wrote is not always clean. Some of the gap TRM found may be classification conservatism rather than evidence agents are absent. Even under that charitable reading, though, the conclusion stands: verifiable autonomous spending is not yet a meaningful share of volume.
Implications for Issuers and Investors
For Circle and Coinbase, x402’s strategic value has been positioning USDC as the settlement asset for machine commerce before that commerce exists. That is a defensible land-grab strategy, but TRM’s data is a reminder that the land is currently mostly empty. Investors pricing “agentic payments” revenue into stablecoin issuer valuations should treat agent-originated volume as a separate line item from protocol volume, and the two are currently being conflated in much of the market commentary.
The finding also gives ammunition to skeptics who argue agentic payments are a bubble. As we have argued before, a bubble in attention is not the same as the rails being wrong — but the burden of proof now sits with the protocols to show agent-initiated volume growing as a share of the total, not just protocol volume growing.
What to Watch
The metric that will settle this debate is not total x402 transaction count but the ratio of payments initiated without human involvement. Expect TRM Labs and competing analytics firms to publish follow-up measurements, and expect x402 supporters to push back on attribution methodology. If agent-share of volume is rising quarter over quarter, the thesis is intact and merely early. If it stays flat while headline volume grows, the growth is coming from ordinary developers using a convenient payment API — which is still a business, just not the one being funded.