ActPlane introduces programmable OS‑level policy enforcement for AI agent harnesses

ActPlane is a new framework that enforces OS‑level policies on AI agent harnesses. It provides a programmable interface for defining permissible actions of agents. The system intercepts system calls and applies constraints based on policy rules. Researchers demonstrate how ActPlane can prevent unauthorized resource access. The approach integrates with existing operating‑system kernels without major modifications. Experiments show reduced risk of harmful behavior while maintaining agent performance. The paper outlines use cases such as sandboxed deployment and compliance monitoring. Future work aims to extend support to diverse platforms and richer policy languages.