For the better part of a year, every announcement covered on this site — AWS AgentCore Payments, Circle's Agent Stack, The Graph's x402 gateway, MoonPay's MoonAgents Card — has rested on a tacit assumption: that AI agents transacting in stablecoins is a real, measurable behaviour, not a deck slide. This week, the first serious data set arrived to test that assumption. The numbers are bigger than I expected, and the structural risk underneath them is more uncomfortable than the headlines suggest.
The new Keyrock report, "Who Pays the Agent?", produced with Coinbase, Tempo, and Virtuals, puts hard figures on the machine economy for the first time. Between May 2025 and April 2026, AI agents processed roughly 176 million on-chain transactions worth $73 million, at an average size of $0.31–$0.48. By the end of Q1 2026, more than 104,000 AI agents had registered across the major protocols. And 98.6% of every cent of that settlement flowed through a single stablecoin: USDC.
The first half of that story is the validation everyone in agentic commerce has been waiting for. The second half is the risk that, in Keyrock's own words, "nobody in the space is publicly discussing — we think they should be."
The Micropayment Thesis Survives Contact with Reality
The most important data point in the report is the size distribution. Roughly 76% of all agent transactions fell below Visa's $0.30 fixed fee threshold. That single statistic does more to settle the "are stablecoins really an agentic-payment primitive" debate than any keynote could. A clear majority of the activity is economically impossible on traditional card rails. The unit economics simply do not exist for a Visa or Mastercard tap on a 9-cent API call, even before considering reconciliation overhead.
By contrast, a USDC transfer on Base costs roughly $0.0001 — about 0.03% of a $0.31 transaction. The micropayment thesis was always credible in theory. Now there is a year of production data showing it works at scale: 176 million transactions, average ticket size in the cents, almost all of it on settlement infrastructure that traditional payments cannot price.
For any payment developer designing API or MCP-server billing, this number is the new floor under the conversation. You no longer have to defend whether per-call stablecoin pricing makes sense. You have to defend why your endpoint isn't offering it.
Why 98.6% USDC Is a Problem, Not an Achievement
Now the uncomfortable part. The same data set that vindicates the rail also reveals how dangerously thin its foundations are. 98.6% in a single stablecoin from a single issuer is not diversification expressing a free-market preference. It is concentration at the scale of a systemic risk.
Read the full article on tomcn.uk →
About the Author
I'm Tom Wang, an AI Developer & Fintech Developer — building AI agents, crypto payment infrastructure, and cross-border payout systems with Rust, Go, and TypeScript. Based in London, UK.
Currently open to new opportunities in fintech, crypto payments, and AI agent engineering.
Top comments (1)
The shift you're flagging, that agent stablecoin transactions are now a measurable behaviour and not a deck slide, is the part most of the agent-payments hype skips straight past, so grounding the concentration argument in the first real dataset is genuinely valuable. The concentration risk is the right thing to worry about precisely because agents amplify it: a human treasurer diversifies out of nervousness, but an autonomous agent will happily route 100% through whatever rail its config defaults to, with no instinct that single-stablecoin dependence is a systemic exposure. So the concentration isn't just a market fact, it's a property of how agents are built, they optimize the path they were given and don't hedge unless told to. That argues for the same discipline I keep coming back to on agent money: the policy layer should encode constraints the agent won't reason its way to on its own, including diversification and exposure limits, not just spend caps. Autonomy without an explicit risk policy concentrates by default. That encode-the-risk-constraints-the-agent-won't-infer instinct is core to how I think about agent commerce in Moonshift. Does the data show the concentration is agent-driven specifically, or just inherited from USDC's overall stablecoin dominance?