“An experiment at Anthropic demonstrated that Claude agents were effective at understanding and acting on user preferences, even when those preferences were not fully articulated.”