AI agents moved $50 billion. Half of it was tests.
Two hundred million payments. Sixty-nine thousand agents doing the buying. Around $50 billion in volume — on a payment code that sat unused in the web since the 1990s.
It is the most impressive number in technology this month, until you reach the footnote: roughly half of it was practice.
Both halves matter, and most coverage prints only the first. What is real: the rails exist, nobody owns them — the Linux Foundation took the protocol over in April, after Coinbase gave it away — and software can now buy things without a human typing in a card number.
What is not real yet: demand. Around half that volume looks like testing rather than trade.
Anyone quoting $50 billion without the asterisk is selling you something.
Infrastructure has arrived well ahead of the use case. Normal for a payment network. Dangerous only if mistaken for traction.
Anyone deciding whether agent payments are a 2026 problem or a 2029 one. On this evidence: build now, forecast later.
An honest split of tests versus real trade. Until someone publishes one, treat $50 billion as capacity, not demand.
The first serious loss in agent payments will not be a hack. It will be a program doing exactly what it was told, at a scale nobody sized. And the fix will not be clever cryptography — it will be a spending limit, which is what humans invented the first time we handed somebody else a card.
Sources: Blockchain.News; AMINA Bank research; Fireblocks agentic finance report.