Agents Act. Systems Remember. Rules Authorize. People Govern.

As AI agents become capable of carrying out consequential work, the central governance problem is not simply what agents can do, but how their actions become trusted facts, how those facts change what is allowed, and where human judgment remains necessary. This paper proposes a four-part architecture: agents act; trusted systems record what happened; rules decide what is now allowed; and people govern when judgment, responsibility, exceptions, or risk require human authority. The architecture separates action, evidence, authorization, and governance rather than collapsing them into a single agentic loop. A central integrity rule follows: the principal receiving new authority must not be able to manufacture all the evidence sufficient to receive it. The paper develops a state-based process model for applying this principle at consequential boundaries while avoiding the opposite extreme of making every agent action a human approval task. The result is a practical architecture for increasing agent autonomy without giving up provenance, accountability, scoped authority, or meaningful human governance.