A Multi-Agent Framework for Large Language Model-Driven SQL Analysis
A Multi-Agent Framework for Large Language Model-Driven SQL Analysis

Summary/Abstract: As organizations accumulate increasingly vast and complex datasets, the need for accessible, intuitive data analysis interfaces has grown. This paper proposes a multi-agent framework that leverages large language models (LLMs) to autonomously explore and analyze SQL databases. Our approach incorporates three specialized agents—Manager, Supervisor, and Database Analyst—each tasked with distinct responsibilities to collectively orchestrate queries, interpret results, and formulate next steps. By breaking down the problem-solving workflow into iterative steps, this framework capitalizes on LLMs’ ability to reason in dialogue-like patterns and generate safe, context-aware SQL queries. We demonstrate the efficacy of our system through an example case study, in which the multi-agent setup retrieves clues from the database, refines queries based on intermediate results, and converges on a solution only with initial human guidance. Our findings highlight the value of role-specific prompts and iterative reasoning loops for mitigating LLM hallucinations and improving query accuracy. While significant challenges remain—such as managing large-scale databases efficiently and ensuring reliable stopping criteria—this research underscores the promise of employing multi-agent LLM systems for flexible, transparent, and robust database exploration.

• Details
• Contents