eDwards: Multi-agent system implementation

Abstract

This paper presents a multi-agent system implementation capable of assisting users with various tasks and processes. It explores previous approaches to conversational agents with the goal of streamlining interactions with processes and reducing friction with involved systems through conversational user experiences (CUX). With the advent of pre-trained Large Language Models (LLMs), new techniques and opportunities have emerged for creating these experiences, accelerating the development of specialized multi-agent systems. This paper examines the technical requirements, including the ability to handle both structured and unstructured knowledge, entity recognition, and the transformation of web forms into dialogues. It also explores different frameworks and the state of the art. Finally, it presents an architecture that leverages three agents to solve different problems, discussing challenges and lessons learned from the implementation of various conversational agent architectures.