Deontic Policies: Governing Agentic AI at Runtime

When AI agents start autonomously calling APIs, installing software, and collaborating across organizations, traditional access control mechanisms quickly fall short. We need a more nuanced governance structure — one that not only dictates what an AI can and cannot do, but also specifies what *must* happen after certain actions (like notifying a security officer) and under what conditions an obligation can be waived. This is precisely the problem tackled by a new paper on arXiv, proposing a framework for agentic AI governance.

The Gaps in Today's Policy Engines

Existing policy languages like XACML, Rego, and Cedar were never designed with the complexities of AI agents in mind. They excel at binary choices — permit or deny — but struggle with obligation rules that demand actions like, "After completing A, B must be executed within 10 minutes." Even more challenging, when two policies conflict (e.g., one requiring notification, another mandating secrecy), these systems lack built-in meta-policy conflict resolution mechanisms. The paper argues that for enterprises to truly control agentic AI, a comprehensive set of norms covering permissions, obligations, dispensations, and priority judgments is essential.

Deontic Logic Makes a Comeback

The research team turned to an ancient yet highly relevant field: Deontic Logic, which specifically studies the relationships between obligations, permissions, and prohibitions. They've extended this into a framework for runtime governance policies, built around four core dimensions:

• Permission/Prohibition: Defines whether an agent can perform an action, aligning with existing policy engines.
• Obligation Lifecycle: Manages the complete state of an obligation, from triggering and activation to fulfillment or timeout.
• Dispensation: Allows for the revocation of an obligation under specific conditions, while ensuring compliance auditing.
• Meta-Policy Conflict Resolution: Automatically arbitrates when rules conflict, based on predefined priorities or contextual factors.

This means that when an AI agent performs a sensitive operation, the system doesn't just log it; it can actively trigger subsequent processes — perhaps automatically generating a report, awaiting approval, or even rolling back changes.

Real-World Impact: Ensuring Enterprise AI Compliance

For enterprises deploying LLM Agents, the practical value of this paper lies in its provision of a deployable governance model. Consider the financial sector, where an AI agent executing a trade might be bound by a "two-person review" obligation. Or in healthcare, accessing patient data could immediately trigger an audit log generation and notification to the data protection officer. These scenarios are difficult to implement elegantly with traditional policy engines, but a deontic logic-based framework offers native support.

Another critical use case is cross-organizational collaboration. When AI agents from different companies interact, their respective policies might clash. The paper's meta-policy mechanism allows for defining "trust but verify" rules — for instance, accepting the other party's obligations but appending local notification requirements.

A Starting Point, Not the Finish Line

The research team openly acknowledges that this framework is currently more of a theoretical model than a production-ready implementation. However, its direction is crystal clear: governance for Agentic AI cannot rely solely on API gateways or firewalls; it must delve into the business logic layer. For developers, a few key takeaways emerge:

• Evaluate existing policy engines: If you're using Rego or Cedar to manage AI agents, check if they support obligations and dispensations. If not, consider extensions or alternatives.
• Monitor standardization efforts: This paper could very well influence the next generation of policy language standards, similar to XACML. Keeping an eye on these developments is wise.
• Start with simple obligations: Even with a complex framework, begin by implementing basic obligations like "notify after operation" in critical processes to build experience.

AI agents are transitioning from experimental tools to production systems, and governance is that often-overlooked yet crucial component. This paper serves as a roadmap, reminding us that security isn't just about controlling permissions; it's about managing behavior and responsibility.