Multi-Agent System for Automated Code Reviews

Abstract

Code reviews play a key role in software quality assurance, yet the process is often time-consuming, inconsistent, and dependent on reviewer expertise. Although recent Large Language Model (LLM) advances have introduced automated review tools, single-model systems often suffer from inconsistency, redundancy, and limited trust when relying on a single agent. This thesis addresses these limitations by proposing a multi-agent LLM-based automated code review system in which four domain-specialized agents, Readability, Refactoring, Performance, and Security, independently evaluate code changes, and collaborate via a Consensus Agent that consolidates, scores, and ranks the issues identified. The system introduces an Evaluation Matrix Consensus Scoring (EMCS) model that balances reliability and impact by combining consensus metrics such as confidence and inter-agent agreement with priority metrics such as severity and category importance to compute a final score for each issue.

Empirical results indicate that the multi-agent approach broadens coverage across categories, reduces redundancy, and yields more stable confidence distributions compared to a single-LLM baseline. The agents exhibit strong domain-specific specialization, with meaningful differences in issue types, severity profiles, and confidence levels, while achieving moderate overlap that supports cross-agent validation. The Consensus Agent significantly improves clarity and actionability by merging semantically similar issues, filtering low-value detections, and integrating reliability and impact metrics. The final ranked output is more concise, structured, and trustworthy than that of individual agents.

This work presents a novel architecture for automated code review that combines domain-specialized agents with a conscientious consensus mechanism. It demonstrates how multi-agent collaboration enhances the reliability, clarity, and usefulness of LLM-generated feedback. Together, these advances demonstrate that multi-agent consensus can improve the trustworthiness and clarity of automated code review, offering a practical pathway for integrating LLMs into developer workflows and opening doors for future research on adaptive weighting and protocol-driven orchestration.