The question
How should an AI workflow decide whether an action is authorized, and how can that decision remain inspectable?
My contribution
I am developing AgentOS around the engineering controls needed to govern agent behavior. The work brings together workflow design, identity, permissions, budget limits, and retained execution evidence.
What I built
The implemented foundation includes a registry of roles and capabilities, action classification, policy checks, scoped approval records, and typed workflow dispatch. Human approvals are bound to the specific proposed action.
Design choices
Authorization is evaluated outside the language model. Approval lifetime and scope are explicit. Execution history preserves the relationship between the request, decision, and resulting action.
Current stage
AgentOS is an engineering prototype tested locally. The next step is to evaluate a complete workflow in a controlled environment before wider use.