Detecção automatizada de code smells em Python utilizando LLMs multiagentes

Abstract

Code smell detection is an essential activity for ensuring software quality and maintainability. Traditional static analysis tools, although widely used, have limitations such as high false positive rates and inability to consider project-specific context. This work investigates the application of multi-agent systems based on Large Language Models (LLMs) for automated detection of implementation code smells in Python projects. The proposed system uses an architecture with a Supervisor agent coordinating eleven specialized agents, each dedicated to detecting a specific type of smell: Long Method, ComplexMethod, ComplexConditional, LongParameter List, Long Statement, Long Identifier, Magic Number, Empty Catch Block, Missing Default, Long Lambda Function, and Long Message Chain. Three language models were evaluated (Claude Sonnet 4.5, GPT-4o-mini, and DeepSeek V3.2) and two types of prompts were compared: elaborate and simple. Experiments were conducted on a dataset of 20 Python files totaling 8,463 lines of code, validated against a reference standard with 411 smell instances. Results demonstrate that the system achieved F1-Score between 61.62% and 65.12%, with GPT-4o-mini presenting the best balance between quality and cost. The analysis revealed that elaborate prompts provide a 4.33 percentage point gain in F1-Score, although the difference is not statistically significant. As an additional contribution, a Visual Studio Code extension was developed that integrates smell detection into the developer workflow.