Who acts when the machine decides? Intentional action and agentic AI

Action theory offers well-developed treatments of tool use and of delegation, but agentic AI belongs to neither category. The user specifies an end; the system autonomously selects the means. What, then, secures the connection between the user’s intention and outcome that distinguishes intentional action from lucky coincidence when things go right? I examine five candidate approaches – Davidsonian minimalism, principal–agent decomposition, strong practical-knowledge views, permissive intention-plus-success accounts, and extended-agency frameworks – and argue that each draws the line in the wrong place, variously under-crediting supervised use, over-demanding means-level transparency, erasing the competence/luck distinction, or inheriting the bloat and glue problems that plague extended-mind theories. The positive proposal then relocates non-deviance from route-identity to governance-stability. On the know-how account defended here, a user intentionally ϕ-s through agentic AI when her success manifests the exercise of supervisory know-how – viz., the competence to define an “acceptable corridor” of outcomes and means, to monitor for departures from that corridor, and to intervene or abort when drift threatens. This framework tolerates the opacity and route-variability characteristic of agentic AI, explains graded attributions across interaction profiles (in/on/out-of-the-loop), and remains neutral on whether the AI itself qualifies as an agent. It thereby vindicates, under specifiable conditions, the ordinary human claim “I ϕ-d through an AI agent” as tracking creditworthy governance.