source: arxiv artificial intelligence: deontic policies for runtime governance of agentic ai systems
level: research
large language model agents that can use tools, change data, and coordinate across organizations need more than access control. they need full enterprise governance that says what agents must do after certain actions, when rules can be waived, and which policies win in a conflict. current policy engines like xacml, rego, and cedar only handle permit and prohibit decisions. they lack obligation lifecycle management, meta-policy conflict resolution, and dispensations that waive standing rules.
the paper proposes a deontic policy model that adds obligations, permissions, prohibitions, and dispensations. it includes a runtime monitor that tracks obligation fulfillment and enforces meta-policies for conflict resolution. this lets organizations define rules like "an agent must notify the ciso after accessing sensitive data" and handle cases where that obligation can be waived under specific conditions. the model also supports precedence rules when multiple policies apply.
the approach is designed for agentic ai systems where llms make sequences of tool calls. the runtime governance checks each action against the policy state, updating obligations and enforcing constraints in real time. this moves beyond static access control to dynamic, context-aware governance that can adapt as agents operate across boundaries. the framework aims to provide the missing pieces for safe deployment of autonomous agents in enterprise settings.
why it matters: as llm agents gain more autonomy, organizations need practical ways to enforce complex rules beyond simple access control, reducing risk of unintended actions.
source: arxiv artificial intelligence: deontic policies for runtime governance of agentic ai systems