An autonomy level says how independent an AI is; an authority contract says exactly where the edges are. Contracts are the enforcement object - not guidance the AI reads, but rules the platform applies to every action before it executes.
What a contract contains
- ›Scope - global, one module, or one agent.
- ›Autonomy level - the ladder position for this scope.
- ›Allowed and denied verbs - the specific actions (create_task, update_deal_stage, send_message...) this scope may or may never perform. Deny wins.
- ›Risk ceiling - the maximum risk class executable without approval.
- ›Monetary ceiling - a hard cap on financial impact per action.
- ›Approver roles - who can approve what exceeds the contract.
- ›Quiet hours - windows when nothing executes (payroll week, launch day).
- ›Rate limit - maximum executions per hour, which also caps credit spend.
Where to set them
Command Center, Autonomy is the control room: every scope, its contract, and recent enforcement decisions. Per-agent overrides let one trusted agent exceed the module default - or one experimental agent run tighter.
How enforcement feels
When an AI attempts something outside its contract, the action is refused or routed to approvals, and the refusal is logged with the reason. Nothing silently fails; nothing silently succeeds.
Tip: Write contracts around verbs, not vibes. "May create and update tasks, may not delete anything, may not touch Finance, 10 actions per hour" is a contract. "Be careful" is not.