Deontic Policies for Runtime Governance of Agentic AI Systems

Researchers propose AgenticRei, a deontic policy framework for governing autonomous AI agents that goes beyond traditional access control by implementing obligations, dispensations, and conflict resolution. The system addresses critical gaps in existing policy engines like XACML and Cedar, enabling enterprises to enforce comprehensive governance constraints over LLM-driven agents that invoke tools, manipulate data, and coordinate across organizational boundaries.

The emergence of autonomous agentic AI systems creates unprecedented governance challenges that existing enterprise policy frameworks cannot adequately address. Traditional access control systems operate on permit/prohibit binary logic, but autonomous agents require governance structures that specify obligations (what agents must do after certain actions), dispensations (exceptions to standing rules), and hierarchical policy conflict resolution. AgenticRei bridges this gap by implementing deontic logic—a formal system for reasoning about obligations, permissions, and prohibitions—expressed through OWL ontologies and evaluated by a dedicated logic engine operating independently from the LLM itself.

The architecture reflects a fundamental architectural principle: governance enforcement should exist outside the model generating the decisions. By decoupling policy evaluation from the LLM, the system prevents agents from manipulating or circumventing governance constraints through prompt engineering or emergent behaviors. This design is particularly critical for high-stakes domains like healthcare, financial services, and cybersecurity where regulatory compliance and liability frameworks demand auditable, deterministic policy enforcement.

The framework's ability to reason over domain-specific ontologies—such as healthcare class hierarchies or data classification schemes—enables policies to adapt dynamically to organizational context without hardcoding every edge case. The proposed integration with industry-standard A2AS (Agent-to-Agent Security) frameworks suggests potential pathways toward standardized agent governance across enterprise ecosystems.

For enterprises deploying autonomous agents at scale, AgenticRei represents a critical infrastructure component rather than a competitive differentiator. The research identifies genuine capability gaps in production systems, but the practical adoption timeline depends on enterprise tooling maturation and regulatory clarity around agent accountability.

• →Existing policy engines (XACML, Rego, Cedar) lack obligation lifecycle management and conflict resolution needed for autonomous agent governance
• →AgenticRei uses deontic logic and OWL ontologies to express permits, prohibits, obligations, and dispensations in a unified framework
• →Policy enforcement operates outside the LLM entirely, preventing agents from circumventing governance through model manipulation
• →The system enables ontological reasoning over domain-specific hierarchies for context-aware policy application in regulated industries
• →Framework composition with A2AS suggests standardization potential for enterprise agent governance architectures