“AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation” was posted to arXiv on August 16, 2023 by a team of fourteen authors led by Qingyun Wu and Chi Wang, with work from Microsoft Research and academic collaborators. It introduced an open-source framework for building applications as conversations among multiple agents, each of which can be backed by a language model, a human, a tool, or some combination of the three.
The central abstraction is the conversable agent. Developers define agents and the patterns by which they talk to one another, using a mix of natural language and code, and the framework handles the message passing that drives a task forward. Because a “human” can be one of the agents, AutoGen unifies fully automated and human-in-the-loop workflows under the same conversational model. The paper demonstrated the framework across mathematics, coding, question answering, operations research, online decision-making, and entertainment applications, arguing that the same multi-agent conversation infrastructure could express applications of very different complexity.
AutoGen became one of the most widely used multi-agent frameworks and seeded a research line at Microsoft that later produced systems like Magentic-One. It is often discussed alongside CAMEL, MetaGPT, and ChatDev as part of the 2023 wave that established “several LLM agents in conversation” as a practical way to build software, rather than a purely academic idea.