Formal Policy Enforcement for Real-World Agentic Systems

ArXi:2602.16708v3 Announce Type: replace-cross Security policy enforcement in contemporary agentic systems predominantly consists of embedding natural-language policies within an agent's system prompt and delegating compliance to the agent's reasoning. This approach admits no formal enforcement guarantee and cannot express policies whose satisfaction depends on the causal history of an execution, a gap that becomes acute in multi-agent systems, where enforcement must reason across agents.