Autonomous AI agents powered by large language models present an unprecedented challenge: systems that can execute code, access databases, install software, and coordinate with other agents across organizational boundaries. A new paper on arXiv proposes deontic policies—a formal governance framework that operates at runtime rather than relying solely on pre-training safety measures. Unlike traditional alignment approaches that embed values during model training, deontic policies function as executable constraints that actively monitor and restrict agent actions in real time. This distinction matters because even well-trained models can behave unpredictably when given access to powerful tools and real-world systems.
The deontic approach treats agent governance as a permission and obligation problem: specifying what actions agents can and cannot perform, under what conditions, and what consequences follow. For example, a deontic policy might permit an agent to query a customer database but prohibit exfiltrating personal information, or allow software installation only after security scanning. These constraints operate independently of the model's internal reasoning, creating a technical safeguard that doesn't degrade agent capability. This is critical for enterprise and financial applications where both autonomous decision-making and strict compliance requirements coexist.
The framework addresses a growing gap in AI safety: as agents become more capable and autonomous, pre-training-based safety measures alone prove insufficient. Runtime governance offers a complementary layer—one that can be updated, audited, and enforced without retraining models. The approach resonates particularly in regulated sectors like finance and healthcare, where supervisory oversight must adapt to emerging risks. By formalizing permissions and obligations through deontic logic, the research provides concrete tools for organizations deploying agentic systems at scale while maintaining human oversight and regulatory compliance.