AI agents spend. You sign off.
Metal holds each agent's on-chain budget, Airwallex holds each card's limit, Purser's code holds the rest; the CFO approves anything beyond them.
The agents wait for the world to change.
Three moments from the run
Stockline raises its price to 0.04 → one purchase is let through on purpose, and Metal refuses it on chain → Treasury replans → two signatures → the mandate is replaced.
Watch it happenA simulated customer disputes 12 orders, saying its agent never placed them. Every order carries its agent's signature, so that claim fails. Three orders had wrong data; the Support agent refunds those 3 (0.15 test USD) itself. "Refund the other 9" goes to the CFO, who rejects it with Face ID.
Watch it happenEarnings swept after finality → euros for the shortfall bought by FX → licence paid by SWIFT after Face ID → a card buy signed with Visa TAP.
Watch it happenProof, from the main run
Replay of the main run, 3 Nov 11:16–11:33 UTC · test money · real Claude decisionsEvery time an agent decided something, it was a Claude call. 0 fell back to code.
Open proof 87payments inside the limits41 sales + 41 supplier purchases + 3 refunds + 1 sweep + 1 card buy, with no approval at the time; the sales are the simulated customer's agent buying from us.
Open proof 16 stopsnothing paid over a limit16 stops, nothing paid over a limit: Claude skipped 12 overpriced checks, Metal refused 2 purchases, the CFO rejected 1 proposal, and 1 planted instruction changed nothing.
Open proof 5Face ID decisionsThe CFO signed each one on his phone; the mandate change needed a second signature too.
Open proof $0.0132Claude cost per decisionWhat the agents' Claude calls cost per decision, at Anthropic's list prices; each decision may be several calls.
Open proofAirwallex declined EUR 45.00 over the card limit (LIMIT_EXCEEDED), separate TAP run, 2 November; not counted in the figures above.