Three-Layer Probabilistic Architecture Is Structurally Required for Safe LLM Agent Deployment
arXiv·medium signal
Position paper argues that enforcing LLM agent safety within a single abstraction layer is not merely suboptimal but categorically insufficient — a structural consequence of how agent execution works. Proposes a three-layer probabilistic assume-guarantee architecture where each layer provides independent safety guarantees. The argument is that no amount of prompt engineering or fine-tuning can substitute for architectural separation of safety concerns in deployed agents.