What question did this study set out to answer?

This work aims to compare different multi-agent architectures for automating Python code refactoring using LLMs.

May 13, 2026Open Access

A Factorial Comparison of LLM-Based Multi-Agent Architectures for Automated Python Refactoring

Key Points

This work aims to compare different multi-agent architectures for automating Python code refactoring using LLMs.
Implemented four LLM-based multi-agent architectures with varying communication structures and coordination strategies.
Utilized a curated dataset of 2,000 Python files from real-world ML projects sourced from GitHub.
Established reproducibility with provided source code, configuration files, and scripts.
Demonstrated effectiveness of dynamic coordination strategies over static in Python code refactoring tasks.
Found no significant difference in performance between hierarchical and centralized communication structures.

Abstract

This repository contains the companion artifact for the paper "A Factorial Comparison of LLM-Based Multi-Agent Architectures for Automated Python Refactoring", submitted to the International Symposium on Empirical Software Engineering and Measurement (ESEM 2026). It includes the implementation of four LLM-based multi-agent system (MAS) architectures that automate Python code refactoring by varying two design dimensions — communication structure (hierarchical vs. centralized) and coordination strategy (static vs. dynamic) — all orchestrated with LangGraph. The artifact provides the full source code, configuration files, a curated dataset of 2,000 real-world ML project Python files sourced from GitHub, and scripts to reproduce the experimental results for RQ1–RQ3. It enables reproducibility, transparency, and future research on collaborative LLM-driven software maintenance approaches.

Read Full Paperexternally

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper