Where should enforcement of a rule like "never transfer more than $10,000 without a second approval" primarily live in a production agent architecture?

Answer: B — In a pre-tool-call hook that checks the amount and blocks or requires approval before execution

A pre-tool-call hook enforces the rule structurally, independent of the model's reasoning. A prompt instruction is advisory and can be overridden by unusual input or a long conversation; an after-the-fact error message doesn't prevent the transfer.