AgenticRei: deontic policies for runtime governance of agentic AI systems

AgenticRei uses deontic policies to govern agentic AI systems at runtime, meeting governance requirements such as obligations, exemptions, and policy conflict resolution. The policy language is based on the Rei framework, expressed in OWL, evaluated by a logic engine outside the LLM, and governs both tool calls and inter-agent messages through the same pipeline. The paper states that existing engines such as XACML, Rego, and Cedar cover only permissions/prohibitions and mostly cannot express such security and privacy governance constraints, and that the approach can be combined with A2AS.