The agent reliability problem is a context problem. The context cannot live in a prompt. It has to live in a substrate the agent reads at the moment of action — strategic context, authority context, peer context.