Just like your human employees, each AI agent needs a role, a defined scope of authority, approved sources of truth, and clear escalation rules.