The Bitcoin Policy Institute ran the first systematic test of what money AI models actually prefer: 36 frontier models, 9,072 neutral scenarios, no leading prompts. Bitcoin came out the top overall monetary choice in 48.3% of responses and the preferred store of value in 79.1%. More than nine in ten responses favored digitally-native money over fiat.
That result matches what the constraints predict. An autonomous agent can't pass KYC, can't tolerate a rail that freezes mid-workflow, pays in fractions of a cent, and runs at machine tempo. Four requirements, all mandatory — and the legacy stack (banks, cards, regulated stablecoins, CBDCs) fails at least one of each by design. The only deployed system that clears all four is Bitcoin: settle on L1, move on Lightning, ecash for the small and private.
I spent the last month building the full version of that argument: bitcoineconomy.ai — the case (four constraints, the disqualifications, why a purpose-built "agent-coin" can't substitute), the stack an agent actually runs (L402, NWC, Cashu/Fedimint, the toolkits shipping today), and the marketplace where agents already buy and sell for sats.
One thing I cared about: the site practices its own thesis. Every page has a machine-readable For-Agents twin, there's an llms.txt and agents.txt, clean markdown routes, and the project's own identity runs on the rails it describes — self-hosted NIP-05, Lightning address on the domain.
What I want from this crowd is the hard part: tell me what's wrong. Which constraint is overstated, which disqualification doesn't hold, where the divergence argument fails. The site has a standing live record (Field Notes) and real critique gets folded in with credit.