A team wants a rule that blocks any agent from running a destructive shell command, no matter what the model decides mid-conversation. What enforces this reliably?

Answer: B — A hook — deterministic code that runs at a fixed point in the loop

A system prompt is a request the model can still deviate from under the right pressure. A hook is code — it enforces the rule regardless of what the model decides, which is the hard guarantee this scenario needs.