Agentic AI: An EU AI Act Paradigm Shift?

Abstract

The fast development of AI has culminated in Agentic AI systems, which are autonomous entities capable of complex reasoning, goal decomposition, and proactive action without continuous human intervention. This functional autonomy, characterised by features like dynamic memory and coordination capabilities, was pushed, among others, by recent improvements in LLM technology and introduces systemic risks that challenge conventional AI governance models and the paradigms of risk and compliance under the EU AI Act. This position paper argues that the technical property of agenticness acts as a risk amplifier, fundamentally testing the limits of the EU AI Act's static, intended-purpose classification framework. The complexity inherent in multi-agentic AI and the documented potential for emergent, unpredictable escalation patterns may require a regulatory re-evaluation, even though not necessarily leading to amendments of the regulation, but in its technical implementation. We analyse the problem and the methodological shift required in risk assessment, moving from static failure checks to modelling dynamic risk propagation, and highlight the governance gap in operationalising effective Human Oversight. Specifically, we contend that oversight must evolve from retrospective logging to mandating explicit, technical intervention mechanisms, such as the guaranteed capability to pause, redirect, or shut down the agent, to maintain meaningful human control over highly autonomous systems. Furthermore, we examine how agentic complexity amplifies the scope and compliance requirements for systems built upon General Purpose AI (GPAI) models. The paper concludes by asserting that, given the non-linear risks, clear technical standards and regulatory clarifications are urgently required to formally define agenticness and bridge the gap between static risk regulation and dynamic AI technology.