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

Qingyun Wu,Gagan Bansal,Jieyu Zhang,Yiran Wu,Beibin Li,Erkang Zhu,Li Jiang,Xiaoyun Zhang,Shaokun Zhang,Jiale Liu,Ahmed Hassan Awadallah,Ryen W White,Doug Burger,Chi Wang

2023-10-04

Abstract:AutoGen is an open-source framework that allows developers to build LLM applications via multiple agents that can converse with each other to accomplish tasks. AutoGen agents are customizable, conversable, and can operate in various modes that employ combinations of LLMs, human inputs, and tools. Using AutoGen, developers can also flexibly define agent interaction behaviors. Both natural language and computer code can be used to program flexible conversation patterns for different applications. AutoGen serves as a generic infrastructure to build diverse applications of various complexities and LLM capacities. Empirical studies demonstrate the effectiveness of the framework in many example applications, with domains ranging from mathematics, coding, question answering, operations research, online decision-making, entertainment, etc.

Artificial Intelligence,Computation and Language

What problem does this paper attempt to address?

The main objective of this paper is to propose a general framework, AutoGen, for constructing applications based on large language models (LLMs) through multi-agent dialogues. Specifically, AutoGen aims to address the following issues:

1. **How to leverage the collaboration of multiple agents to extend the capabilities of LLM-based applications?** The paper points out that as the application scope of LLMs in solving real-world tasks continues to expand and the complexity of tasks increases, using multiple agents for collaboration can effectively extend the capabilities of a single agent.

2. **How to simplify the development process of LLM-based applications?** To support the development of applications across different domains and complexities, the paper proposes a multi-agent dialogue-based approach to simplify complex development processes and enable developers to easily define the interaction behaviors between agents.

3. **How to design individual agents that can collaborate effectively?** The paper explores how to design agents that possess specific capabilities and can participate in multi-agent collaboration. These agents need to be customizable, reusable, and able to work efficiently in a multi-agent environment.

4. **How to achieve flexible inter-agent dialogue patterns?** To meet the needs of different applications, the paper investigates how to define and control inter-agent dialogue patterns through natural language or programming languages, including single-turn or multi-turn dialogues, different user participation methods, etc.

In summary, the paper attempts to solve the problem of establishing a general framework that allows developers to use multi-agent dialogues to construct a range of LLM-based applications, thereby covering multiple fields from mathematics and coding to online decision-making, and significantly reducing development difficulty.