Autonomous Legacy Web Application Upgrades Using a Multi-Agent System

This project introduces a novel Large Language Model (LLM)-based multi-agent system designed to autonomously upgrade legacy web applications to their latest versions. The proposed system effectively distributes tasks across multiple phases, ensuring all files are updated systematically and efficiently.

To evaluate the multi-agent system, Zero-Shot Learning (ZSL) and One-Shot Learning (OSL) prompts were employed, providing consistent instructions for both approaches. The evaluation process involved updating a set of application view files and measuring the frequency and types of errors generated in the resulting files. For more complex tasks, the number of successfully fulfilled requirements was counted. The prompts were executed both within the multi-agent system and with standalone LLMs to establish a comparative analysis. To account for the stochastic nature of LLMs, the process was repeated multiple times.

The results demonstrate that the multi-agent system effectively maintains context across tasks and agents during the update process. In certain scenarios, it outperformed the standalone LLMs by providing superior solutions. These findings highlight the system's potential as a foundational framework for future implementations of model-based updates to existing codebases.

This dataset serves as a key resource for evaluating the performance of LLM-driven systems in automated code updates and paves the way for further advancements in this domain.