Preliminary Reflections on the Legal Governance of Agentic AI: A Private- and Public-Law Reconstruction through the Permission–Action–Oversight–Liability Framework
Abstract
Artificial intelligence has moved beyond the generative stage—producing outputs and proposing decisions to humans—toward an agentic stage, in which systems autonomously pursue delegated goals and act directly upon external systems. Yet the prevailing ‘risk-based’ approach—the high-risk classification of the EU AI Act and the ‘high-impact AI’ regime of Korea’s AI Framework Act—classifies systems statically by their field of use and focuses on outputs, and thus fails to capture agentic AI, whose risk is dynamic and relational as it acts on delegated authority. This article proposes shifting the locus of regulation from a model’s static risk tier to the dynamic chain of Permission–Action–Oversight–Liability, and develops this framework simultaneously in private and public law through the functional comparative method. In private law, it rejects agency theory in favour of an instrumentality theory grounded in the doctrine of attribution of declarations of intent, supplemented by strict and tiered liability. In public law, it revisits the prohibition of automated discretionary dispositions under Article 20 of the Framework Act on Administration, proposing a conditional exception, a negative list of non-delegable domains, and the reinforcement of the right to explanation and state-liability doctrine. It then normativises a ‘chain of trust’ and ‘operational evidence’ as control infrastructure common to both domains, synthesised under the principle of ‘managed autonomy.’ Finally, through comparison with global frameworks and Asian regulatory pathways, it derives concrete legislative proposals for a distinctive K-AI governance, for a country that has become the earliest full-scale implementer of comprehensive AI regulation.