An AI Agent Shouldn't Be Able to Execute Everything It Proposes
What a model proposes, what a system can execute and what has the authority to be executed are three different things. The difference between them has to be built into the architecture, not just requested in instructions.
A model can propose an action, and a system can have the technical capability to carry it out. But the authority to execute must be granted separately and enforced by a control layer that the model does not control on its own.