How continuity distinguishes autonomy from agency in agentic AI

Recent discourse increasingly describes advanced artificial intelligence systems as ‘agentic’. Planning-capable language models, autonomous workflows, and multi-agent architectures are said to exhibit agency because they pursue goals, initiate actions, and coordinate behaviour over time. This article argues that such characterisations risk conflating operational autonomy with agency in the sense relevant to responsibility, authority, and governance. It accepts thin engineering uses of ‘agentic’ as a label for autonomous task execution, but rejects the normative inference that such systems are agents. Drawing on a continuity-based account, the article argues that agency in the sense relevant to responsibility, authority, and governance is not a behavioural achievement but a structural response to continuity pressure: the need to preserve a unified locus of evaluative authority across incompatible future trajectories. Optimisation, planning, and coordination presuppose such unity; they do not generate it. The account complements gradient or multi-dimensional theories of artificial agency by identifying a structural condition they presuppose but do not supply. Through cases involving LLM tool-use agents, reinforcement-learning systems, multi-agent workflows, and institutional decision systems, the article shows that systems governed by externally specified standards may optimise effectively while lacking authorship of those standards. The principal danger lies in attributing agency where continuity, authorship, and responsibility are structurally absent.