Formal Policy Enforcement for Agentic Systems

The researchers propose a formal policy enforcement method for agentic systems based on runtime monitoring, which can allow or block an Agent's pending actions according to user-specified policies. Users write the intended policies in Datalog, and the system enforces the policies at runtime by constructing an event graph; when the Agent is about to perform an action such as a tool call, the method combines the current system state and historical events to determine whether the action violates the policy, then blocks or allows it accordingly. The authors implemented the method and evaluated it on multiple benchmarks, stating that it demonstrates value in real-world scenarios and viewing it as a step toward provably safe and reliable agentic systems.