Foreman watches vibration on a machining line, projects when a bearing will cross the ISO 10816-3 stop the machine threshold, and buys the part before the line stops. Payment settles into on-chain escrow against a spend permission signed weeks earlier.
Sensors went on the machines. A human was left to read the dashboard, raise a requisition, chase three quotes, wait for a PO number and phone the supplier — while the bearing kept degrading.
The data loop was automated. The decision-to-cash loop was not, and that is the one that costs a shift.
The warning is there. The part is cheap. What is missing is the authority to act on the warning inside the window it gives you.
vibration telemetry ISO 10816-3 zones, log-linear RUL trending
│
▼
maintenance agent 4 tools · reads health, stock, quotes · signs
│
▼
spend permission autonomous ≤ ceiling · human above it · 30-day budget
│
▼
supplier paid on confirmed receipt, against a despatch documentEvery step is a real transaction on Base Sepolia. The reasoning trace streams as it happens rather than arriving after a spinner, because a shift assessment is a sequence of decisions and the sequence is the part worth watching.
$180 bearing, 58 hours of life left, zero on the shelf, a vetted supplier who can deliver in 36. Proposed and funded in one transaction. The line never stops and nobody raised a requisition.
$4,000 spindle is proposed and left there. No Funded event, no money moved, until a separate key approves it — and that approval bypasses the cap, because the cap bounds the agent, not the plant.
This is what “humans handle exceptions, machines handle routine” looks like when it is enforced by a contract instead of written in a policy document.

lib/agent.ts rather than on chain, because the price list is not on chain — and saying so is cheaper than having it noticed.Foreman0x6cc8fafc87328a087ac0da2d0c8cae7f9bec2e9a
Verified, so the bytecode running on Base Sepolia can be checked against the source in the repo rather than taken on trust. Press Run agent on the live deployment and it spends actual testnet money on your behalf — bounded by the same contract a plant would rely on.
Every fixture is a seam with a real implementation behind it, and the fixture stays the default so the public demo still runs offline.
What a pilot still is not: mainnet USDC, and an audited contract. Both are deliberate — get the plant data and the loop right where a mistake costs nothing, then decide whether real money is worth an audit.
The next step is a paid pilot with a single Malaysian precision-machining plant: one machine, real telemetry, real ERP, testnet settlement — to prove the loop where a mistake costs nothing. The integration layer for that is already built and tested.