<p>Recent advances in artificial intelligence, particularly large language models (LLMs), have intensified debates about autonomy, agency, and risk. These debates are often shaped by science-fiction imaginaries that oscillate between trivialization and alarmism, and that can influence regulatory intuition as much as technical evidence. This paper argues that contemporary LLM-based systems are best understood, in isolation, as forms of interface intelligence rather than autonomous agents. To clarify this distinction, it analyzes C-3PO and HAL 9000 not as predictions, but as idealized configurations of artificial agency. C-3PO exemplifies systems characterized by exogenous activation, mediational intelligence, and the absence of persistent goals. HAL 9000 exemplifies a configuration in which goal persistence, operational authority, environmental access, and weak corrigibility transform intelligence into autonomous optimization. Read together, these cases help delineate an agency spectrum that remains relevant for contemporary AI systems. Building on philosophy of action, AI safety research, neurobiology, and sociotechnical theory, the paper argues that the most plausible near- to mid-term societal risks of AI do not require intrinsic machine motivation. They arise from delegated agency: the embedding of powerful, reliable systems within institutional contexts that enable the frictionless execution of human purposes. This implies a governance approach in which agency is evaluated as an architectural and institutional property rather than as a feature of isolated models.</p>