Intelligent Governance in Cyber-Physical Systems

The spread of agentic artificial intelligence has turned digital systems from purely analytical tools into agents that can interpret a goal, design a multi-step plan, select tools and act in digital or physical environments. Connecting such agents to digital twins and Internet of Things equipment delivers capabilities such as real-time decision making and continuous optimisation, but it also raises risks including unintended action, reliance on invalid data, prompt injection, propagation of decision error, absence of auditability and ambiguity of responsibility. The aim of this study is to design an integrated framework for governing agent autonomy in physical-digital environments. The research is conceptual-developmental and was conducted through a purposive review of the scientific literature, an analysis of standards and artificial intelligence governance documents, and a synthesis of system architecture constructs. The main finding is the A-DTIG framework, which comprises seven layers: perception and identity, twinning, agentic cognition, multi-agent coordination, controlled execution, trust and audit, and human governance. Within the framework, an authority gateway evaluates every action against agent identity, mission goal, risk level, data quality, simulation outcome and organisational policy. The digital twin acts as the pre-action test environment, edge computing as the local line of defence, the permissioned ledger as the record of evidence, and the human as the party that sets goals, limits of authority and ultimate responsibility. Authority levels, evaluation indicators, research propositions, a smart-factory scenario and a five-stage maturity model are also proposed. The results indicate that trustworthiness should not be measured at the model level alone but must span the full cycle of perception, decision, authorisation, action and accountability. Sustainable autonomy is not achieved by removing the human; it requires bounded and incremental authority, risk-based control, pre-action simulation, post-action monitoring and reversibility.

کلیدواژه ها:

Agentic artificial intelligence ، Digital twin ، Internet of Things ، AI governance ، Trustworthiness