AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework

Abstract This technical report presents AutoGen,1 a new framework that enables development of LLM applications using multiple agents that can converse with each other to solve tasks. AutoGen agents are customizable, conversable, and seamlessly allow human participation. They can operate in various modes that employ combinations of LLMs, human inputs, and tools. AutoGen’s design offers multiple advantages: a) it gracefully navigates the strong but imperfect generation and reasoning abilities of these LLMs; b) it leverages human understanding and intelligence, while providing valuable automation through conversations between agents; c) it simplifies and unifies the implementation of complex LLM workflows as automated agent chats. We provide many diverse examples of how developers can easily use AutoGen to effectively solve tasks or build applications, ranging from coding, mathematics, operations research, entertainment, online decision-making, question answering, etc.

Introduction. Large Language Models (LLMs), like GPT-4, have demonstrated exceptional capabilities for reasoning, creativity, and deduction in many new AI applications, ranging from robot lawyers [15], tools to assist with fundamental research [7], and a plethora of highly-capable chatbots (e.g., Bard, Bing Chat, ChatGPT+, Copilot). As single LLM calls often lack the ability to reflect, possess no working memory or scratch pad, and cannot act or perceive beyond the knowledge obtained from their training data [8], augmenting language models with tools that facilitate perception and action, such as the plugin that enables ChatGPT+ to utilize fresh web pages from the internet [36], can significantly improve performance. Despite the remarkable success the LLM-with-tools paradigm [41] has enabled, it typically utilizes a single LLM agent. Considering the evolving range of real-world tasks that could benefit from LLMs and the intrinsic weaknesses of using a single agent [29, 16, 8], a promising direction for future LLM applications is to have multiple agents work together to solve complex tasks.

Discussion / Conclusion. The applications in Section 4 show how AutoGen can not only enable new applications but also renovate existing ones. For example, in A1 (scenario 3), A5, and A6, AutoGen enabled creating multi-agent chats that follow a dynamic pattern instead of a fixed back-and-forth. And, in both A5 and A6, humans can participate the activities together with multiple other AI agents in a conversational manner. Similarly A1 (the first two scenarios), A2, A3, and A4 show how popular applications can be renovated quickly with AutoGen. Despite the complexity of these applications (most of them involve more than two agents or dynamic multi-turn agent cooperation), our AutoGen-based implementation remains simple, demonstrating promising opportunities to build creative applications and a large space for innovation. We further reflect on why we observed these benefits from AutoGen in these applications, and we believe there are a few reasons: • Collaborative/adversarial agent interactions: Like many collaborative agent systems (e.g., [14]), agents in AutoGen can share information and knowledge, allowing them to complement each other’s abilities and collectively arrive at better solutions.