Decision Before Action: Rethinking Agentic AI for Risk-Aware Physical Systems

Recent advances in agentic AI have significantly expanded the execution capabilities of autonomous systems operating in the physical world. However, much of this progress implicitly assumes that stronger execution capabilities naturally translate into better decisions. In physical systems operating under uncertainty, this assumption is increasingly fragile. Factors such as environmental uncertainty, asymmetric failure costs, safety considerations, and resource limitations introduce situations in which whether and when to act can be more consequential than how to act. In this paper, we propose a conceptual reframing of agentic AI that explicitly separates decision-level reasoning from execution and planning. We posit that the decision layer governs action commitment, deferral, escalation, and termination based on risk-aware assessment under uncertainty, rather than task performance optimization alone. Through illustrative examples drawn from drones, wearable assistants, and other embodied agents, we highlight recurring failure modes of capability-first designs and show how risk-aware decision abstraction clarifies design trade-offs that are often implicit or overlooked. Finally, we outline open research questions intended to guide future work on robust, human-aligned agentic systems operating in the physical world.