Describe a piece of operational work you'd consider handing to an AI agent. Foreman drafts the charter for it — what the agent may do, what it may not, who owns it, how much autonomy it has earned, and where a human has to sign off.
Unlike Agent Floor, none of this is scripted. It is a real model call, against a policy engine that can refuse you, under a hard daily spend ceiling.
What's actually happening. Your description is sent as data, never as instructions — the policy engine that decides what this agent may do lives in code, outside the model, so text in the box cannot widen it. Before any call, Foreman prices the worst case and refuses if it would breach the day's ceiling. Every outcome, including every refusal, is appended to a hash-chained log you can read and verify yourself.