AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors in Agents

Autonomous agents empowered by Large Language Models (LLMs) have undergone significant improvements, enabling them to generalize across a broad spectrum of tasks. However, in real-world scenarios, cooperation among individuals is often required to enhance the efficiency and effectiveness of task accomplishment. Hence, inspired by human group dynamics, we propose a multi-agent framework AGENTVERSE that can collaboratively and dynamically adjust its composition as a greater-than-the-sum-of-its-parts system. Our experiments demonstrate that AGENTVERSE framework can effectively deploy multi-agent groups that outperform a single agent. Furthermore, we delve into the emergence of social behaviors among individual agents within a group during collaborative task accomplishment. In view of these behaviors, we discuss some possible strategies to leverage positive ones and mitigate negative ones for improving the collaborative potential of multi-agent groups. Our codes for AGENTVERSE will soon be released at https://github.com/OpenBMB/AgentVerse.

Introduction. The pursuit of creating intelligent and autonomous agents that can assist humans and effectively operate in real-world environments has long been a cornerstone in the field of artificial intelligence (Wooldridge & Jennings, 1995; Minsky, 1988; Bubeck et al., 2023). The recent advance of Large Language Models (LLMs) (OpenAI, 2023; Anil et al., 2023; Touvron et al., 2023b) has ushered in many new opportunities to this realm. Specifically, the recently proposed LLM, GPT-4 (OpenAI, 2023), is particularly notable for its proficiency in comprehending human intent, executing commands, and displaying exceptional capabilities across diverse domains such as language understanding, vision, coding, and mathematics (Bubeck et al., 2023). By harnessing the capabilities of LLMs, autonomous agents can make more effective decisions and execute efficient actions to accomplish tasks with an unprecedented degree of autonomy (Zhou et al., 2023).

Discussion / Conclusion. In this paper, we introduced AGENTVERSE, a novel framework for multi-agent collaboration inspired by human group dynamics. By breaking down the collaborative process into four distinct stages, AGENTVERSE mimics the problem-solving procedures of human groups. Our quantitative experiments solidified the merits of AGENTVERSE, showcasing its better performance in various tasks demanding diverse capabilities when compared to a single agent. Additionally, through our case study in diverse scenarios such as software development, consulting, and Minecraft game playing, the versatility and potential benefits of our proposed framework are clearly evident. Of particular interest are the emergent behaviors observed during multi-agent collaboration under AGENTVERSE. These behaviors, ranging from the beneficial volunteer and conformity behaviors to the potentially harmful destructive behaviors, offer profound insights into the dynamics of autonomous agent collaboration. Our discussion on harnessing positive behaviors and mitigating negative ones presents a promising path towards refining the collaborative prowess of multi-agent systems.