Reproduce “ActPlane: Programmable OS-Level Policy Enforcement for Agent Harnesses” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yusheng Zheng, Tianyuan Wu, Quanzhi Fu, Tong Yu, Wenan Mao, Tao Ma, Dan Williams, Wei Wang and 1 more
ActPlane introduces a programmable OS-level policy enforcement system for AI agents, leveraging eBPF to bridge the semantic gap between high-level policies and kernel actions. It allows agents to define contextual rules via a DSL with semantic feedback, demonstrating up to 31 percentage points higher decision compliance than baselines and preventing 74% of unsafe behaviors with minimal overhead.
Key insights
- A significant "semantic gap" exists between high-level natural language policies for AI agents and the concrete, context-dependent system actions required for their enforcement.
- OS-level observability is critical for enforcing most agent policies, especially those involving indirect system actions or cross-event dependencies like data flow and temporal ordering.
- Providing semantic feedback on policy violations, rather than opaque error codes, is essential for AI agents to understand and adapt their behavior for compliance and task recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.