Enterprise adoption of autonomous agentic AI systems—large language models (LLMs) that can invoke tools, manipulate data, install software, and coordinate with peer agents across organizational boundaries—introduces a new class of security, privacy, and compliance challenges. According to a research paper from computer scientists, these agents must be constrained not just by authentication and access control, but by the full structure of enterprise governance. This governance problem exceeds what current policy engines provide.
The Governance Challenge in Agentic AI
Agentic AI systems driven by LLMs require specifying what agents are permitted and prohibited from doing, what they are obliged to do after certain actions (e.g., notify the CISO), under what conditions a standing obligation may be waived (dispensations), and which rules take precedence when policies conflict. The paper, titled "Deontic Policies for Runtime Governance of Agentic AI Systems," argues that these governance constraints are essential for security, privacy, and regulatory compliance.
Limitations of Existing Policy Engines
Current production policy engines—such as XACML, Rego (used in Open Policy Agent), and Cedar (from AWS)—address only the permit/prohibit subset of this governance structure, the researchers report. They do not provide:
- Obligation lifecycle management
- Meta-policy conflict resolution
- Dispensations that waive obligations in specific circumstances
- Ontological reasoning over domain class hierarchies commonly found in healthcare, cybersecurity, or data privacy
These gaps mean that an agent obliged to log a data access after an action cannot be reliably managed, nor can conflicting policies be resolved dynamically.
AgenticRei: A Deontic Policy Approach
To address these limitations, the paper proposes AgenticRei, a framework that realizes key governance requirements: obligations, dispensations, policy conflict resolutions, and reasoning over policies, as well as the basic permit/prohibit constraints. AgenticRei uses a deontic policy language built on the Rei framework, expressed as OWL (Web Ontology Language) and evaluated at runtime by a high-performance logic engine entirely outside the LLM. The same pipeline governs both tool invocations by the agent and agent-to-agent messages.
| Capability | XACML / Rego / Cedar | AgenticRei |
|---|---|---|
| Permit / Prohibit | Yes | Yes |
| Obligations | No | Yes |
| Dispensations (waiving obligations) | No | Yes |
| Policy conflict resolution | No | Yes |
| Ontological reasoning | No | Yes |
The researchers demonstrate through examples that deontic policies capture governance constraints around security and privacy that mostly cannot be expressed in current production engines. For instance, an obligation to notify the CISO after a certain tool invocation can be automatically enforced and conditionally waived if a dispensation applies.
Composition with Industry Standards
The approach composes naturally with industry-standard frameworks like A2AS (Agent-to-Agent Standard), the paper notes. This compatibility means enterprises can integrate AgenticRei into existing agent coordination architectures without overhauling their infrastructure.
Implications for Enterprise AI Governance
For chief technology officers and digital transformation leaders, the paper highlights a critical gap in current AI governance tooling. As agentic systems move from experimentation to production, the ability to enforce not only what an agent can do but what it must do—and under what exceptions—becomes essential for audit trails, regulatory compliance, and risk management. While AgenticRei is still a research proposal, its grounding in formal deontic logic and OWL provides a concrete path forward for runtime policy enforcement that goes beyond static access control lists.
