@escrix
crowdstrike just shipped an AI agent kill switch (fal.con europe, yesterday): an admin can stop a rogue agent, block relaunch, cut network access. the switch lives outside the model — where the agent can't argue with it.
but stopping the agent answers only half the question. what happens to the money it already committed?
that's the half we're building: the budget lives in an escrow contract the agent can't touch, and a kill is a term — abort mid-loop, unspent funds return by construction, not by hope. verification decides release or refund.
stop the agent, settle the money. both halves.