InvAgent: A Large Language Model based Multi-Agent System for Inventory Management in Supply Chains

TL;DR

This work proposes InvAgent, a large language model (LLM)–based zero-shot multi-agent system for inventory management in multi-echelon supply chains. By assigning an LLM-driven agent to each stage and a user proxy to coordinate, InvAgent enables adaptive, explainable decisions through chain-of-thought reasoning and structured prompts without prior training. The framework is evaluated across multiple demand scenarios, showing competitive performance, especially under variability, while highlighting the benefits of prompt design and CoT for transparency and robustness. The findings suggest that LLM-based autonomous agents can reduce inventory costs and stockouts, offering a scalable, explainable alternative to traditional heuristics and reinforcement learning approaches, with avenues for RL fine-tuning and real-world data integration in future work.

Abstract

Supply chain management (SCM) involves coordinating the flow of goods, information, and finances across various entities to deliver products efficiently. Effective inventory management is crucial in today's volatile and uncertain world. Previous research has demonstrated the superiority of heuristic methods and reinforcement learning applications in inventory management. However, the application of large language models (LLMs) as autonomous agents in multi-agent systems for inventory management remains underexplored. This study introduces a novel approach using LLMs to manage multi-agent inventory systems. Leveraging their zero-shot learning capabilities, our model, InvAgent, enhances resilience and improves efficiency across the supply chain network. Our contributions include utilizing LLMs for zero-shot learning to enable adaptive and informed decision-making without prior training, providing explainability and clarity through chain-of-thought, and demonstrating dynamic adaptability to varying demand scenarios while reducing costs and preventing stockouts. Extensive evaluations across different scenarios highlight the efficiency of our model in SCM.

Figures (11)

• Figure 1: The framework of InvAgent, a LLM-based zero-shot multi-agent inventory management system. Firstly, the user proxy resets the environment at the beginning of the first round. Secondly, the user proxy requests the state of the current round for each stage from the environment. Then, the user proxy provides the current state to each stage and requests the action from it. Finally, all agents take actions together and move to the next state.
• Figure 2: The flowchart of multi-echelon supply chain inventory management. Raw materials flow through each stage, comprising inventory storage and manufacturing facilities. The upstream factory at stage $i$ supplies intermediate products to the downstream stage $i-1$, where they are stored as inventory. Stage 0 (retailer) provides final products to satisfy customer demand.
• Figure 3: System messages providing essential information, such as definitions, roles, and goals in the supply chain.
• Figure 4: Prompt provided to LLMs for inventory management simulation. State description, demand description, downstream order description, and strategy description are shown in Figures \ref{['fig:state']}, \ref{['fig:demand']}, \ref{['fig:downstream']}, and \ref{['fig:strategy']}, respectively.
• Figure 5: State descriptions providing the current state for each agent in each period. For the previous sales, we select recent $L_{max}$ periods, and for the arriving deliveries, we select next $L_m$ periods.