guardrail
A rule built around an agent that blocks actions it must never take.
A guardrail is a check that sits outside the model and limits what an agent can do. It might block payments above a limit, or refuse to delete data. Some guardrails are enforced by the software and cannot be talked around. Others are only instructions the model is asked to follow. The enforced kind is the one to rely on.
Think of it as
Like the barrier on a mountain road. It does not steer the car, but it stops the car going over the edge.
Example
A shopping agent can add items and apply coupons. A guardrail in the shop's checkout code refuses any order over ₹20,000 without a person's approval, whatever the agent asks.