“Cognitive Sandwich” architecture for LLM agents in logistics

Traditional decision support systems in supply chain management often lack the flexibility to handle unpredictable disruptions, while emerging autonomous agents powered by large language models (LLMs) struggle with hallucination risks and a lack of mathematically guaranteed plans. To address this challenge, this paper proposes the "Cognitive Sandwich" architecture, a hybrid method that surrounds the LLM's semantic reasoning with strict, deterministic constraints. By utilizing the Model Context Protocol, this approach grounds the agent's decision-making in real-world data from a digital twin and validates it through a mathematical solver. Our experimental results, based on testing across 78 port infrastructure degradation scenarios, confirm the architecture's effectiveness. During these tests, we also identified and classified specific behavioral errors in LLM agents, such as "cognitive escape" (attempting to invent unauthorized network routes) and "Stubborn Agent Syndrome" (falling into infinite loops when attempting to iteratively resolve strict mathematical constraints). From a practical perspective, this verified architecture establishes a strong foundation for building reliable and auditable multi-agent systems in logistics, representing a crucial step toward intelligent, adaptable supply chain orchestration under high uncertainty.