The integration of agentic artificial intelligence (agentic AI) and digital twins (DTs) enables decision-making systems that are intelligent, adaptive, and goal-oriented. This paper advances that convergence through a multilayer integration framework that organizes perception, knowledge and data management, LLM-based reasoning, learning, decision-making, action execution, and feedback adaptation into a cohesive structure. Within this framework, LLM-driven agents act as cognitive cores for context-aware planning and collaboration, while the digital twin provides a shared situational model that maintains real-time state, simulation capabilities, and policy constraints. The contributions are threefold: (i) a novel architecture that operationalizes continuous learning and closed-loop autonomy through tightly coupled agents and a DT; (ii) a use-case demonstration in power balancing for electrical grid management, where agents coordinate demand forecasting, distributed energy resources, and network constraints via a grid digital twin; and (iii) an analysis of key enablers and challenges, such as model synchronization, interpretability, safety, and cognitive load distribution, necessary for scalable and trustworthy deployment. The proposed integration offers a principled path to resilient, transparent, and data-efficient decision-making systems for cyber–physical infrastructure and provides a foundation for extending agentic AI-DT synergies to other safety-critical domains. • Integrates agentic AI with digital twins for adaptive decision-making. • Uses LLMs as cognitive cores for reasoning and goal-oriented autonomy. • Enables continuous learning through real-time feedback and simulation.
Hasan et al. (Sun,) studied this question.