
01
Tool access changes the risk of an assistant
A model that explains a policy can be wrong in a way that a reader notices. A model that changes an account or submits an order can create consequences before anyone reads its explanation. That is why an agent’s authority should be designed independently from the quality of its reasoning.
Think of the model as a proposal generator. It can suggest what should happen, but a policy boundary decides whether the proposed action is permitted for the current user, target, and context. A natural-language instruction should not silently broaden that boundary.