Large Language Model-Driven Data Analysis Agents: A Systematic Review of Architectures and Code Generation Strategies

Autonomous data analysis, once the domain of experienced programmers, is being reshaped by large language models (LLMs) — a shift this paper examines through a structured systematic review. Models like GPT-4, Claude, and LLaMA have reached a level of proficiency in natural language understanding and code generation that now makes it feasible to embed them inside agent-based pipelines capable of planning, executing, and interpreting analytical tasks with little human intervention. To bring conceptual order to this rapidly growing body of work, the review proposes a two-dimensional taxonomy organized around two design choices: system architecture (single-agent versus multi-agent) and code generation modality (Python-centric versus SQL-centric). Within this framework, three lines of research emerge as dominant: systems built around a single execution agent, multi-agent collaborative pipelines, and text-to-SQL generation architectures. Drawing on representative work published at leading venues between 2023 and 2025, each line is assessed with respect to planning strategies, execution feedback loops, and error-correction mechanisms. A cross-cutting analysis finds that how a system handles execution feedback — rather than whether it uses one agent or several — is the factor that most reliably separates stronger from weaker designs. The paper closes by mapping out three underexplored problems: verifying code across modalities, developing unified evaluation benchmarks, and extending multi-agent coordination into SQL-oriented settings.