Constructing a multi-agent system

Abstract

In this thesis, a multi-agent system (MAS) was developed to address the limitations of traditional single-agent systems such as OpenAI's GPT-4. While these models are effective for general tasks, they often face difficulties when tasked with complex, multi-step problems. To overcome these challenges, a system was designed in which multiple specialized agents collaborate, with each agent focusing on different aspects of a task. The backend was constructed using Python, while Vue.js was utilized for the frontend to provide a flexible and user-friendly interface. The multi-agent approach enables tasks to be divided into smaller, more manageable parts, resulting in improved accuracy and efficiency. The architecture allows for the seamless addition of new agents, making the system scalable and adaptable. A comparison between MAS and a single-agent model (ChatGPT) was conducted through a series of tasks, including translation, summarization, and file management. It was demonstrated that MAS handles complex, multi-step tasks more effectively, providing greater precision and advanced functionality. The findings illustrate that a distributed, multi-agent model can more efficiently manage specialized tasks, positioning it as a strong candidate for real-world applications in natural language processing.