Collaborative Multi-Agent LLM Frameworks for Automated Legacy Software Modernization

Abstract

Legacy software modernization is a critical challenge for organizations seeking to improve system maintainability, scalability, security, and compatibility with contemporary technologies while preserving existing business logic. Traditional modernization approaches often require substantial manual effort, extensive domain knowledge, and considerable time to analyze, refactor, migrate, and validate aging software systems. Recent advances in Large Language Models (LLMs) offer promising opportunities to automate these activities; however, relying on a single LLM may be insufficient for managing the complexity, interdependencies, and verification requirements of large-scale legacy systems. This abstract explores a collaborative multi-agent LLM framework for automated legacy software modernization, in which specialized intelligent agents coordinate to perform code analysis, dependency identification, architectural assessment, refactoring, code translation, test generation, and validation. Through structured communication, task decomposition, shared contextual knowledge, and iterative feedback, the framework aims to improve the accuracy, consistency, and efficiency of modernization workflows. A supervisory agent coordinates task execution, while domain-specific agents independently address modernization activities and collaborate to resolve inconsistencies and verify generated transformations. The proposed approach also incorporates human oversight, traceability, and automated testing to mitigate risks associated with semantic errors, hallucinations, and unintended changes in system behavior. The framework is intended to support incremental modernization while preserving functional equivalence and reducing operational disruption. Its effectiveness can be evaluated using metrics such as transformation accuracy, test coverage, behavioral equivalence, execution time, modernization cost, and defect rate, with comparisons against traditional methods and single-agent LLM approaches. By combining agent specialization with collaborative reasoning and systematic validation, this approach provides a potential foundation for more reliable, scalable, and cost-effective legacy software modernization.