“CAMEL: Communicative Agents for Mind Exploration of Large Language Model Society” was posted to arXiv on March 31, 2023 by Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem of KAUST, and was accepted at NeurIPS 2023. It was one of the earliest papers to study what happens when two language model agents are assigned roles and left to cooperate on a task largely on their own.
The core technique is role-playing driven by what the authors call inception prompting. One agent is cast as a user or instructor and another as an assistant, and a carefully constructed initial prompt sets them in motion so they can carry a task forward through conversation without a human steering each turn. The paper presents this as a scalable way to study the cooperative behaviors and capabilities that emerge in a small society of agents, and observes the failure modes - such as role flipping and conversations that drift or never terminate - that the prompting has to guard against. The authors released the framework as open-source software to support further research.
CAMEL helped establish the role-playing pattern that recurs across later multi-agent systems and grew into a broader open-source agent community. It is usually grouped with AutoGen, MetaGPT, and ChatDev as part of the 2023 surge of work on getting multiple LLM agents to collaborate.