When an Agent's capability and blast radius grow together, making the model more aligned is not enough. Anthropic's answer: shift the focus from supervising what the model does to constraining what it can do.