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

In 2023, most LLM applications were a single prompt or a fixed chain. Anything more ambitious, like a model that writes code, runs it, reads the error, fixes it and asks a person when stuck, had to be wired by hand for each task. Qingyun Wu, Chi Wang and colleagues at Microsoft Research, Penn State, the University of Washington and Xidian University built AutoGen, an open-source library that treats such applications as conversations between agents.

The first idea is the conversable agent. Every participant exposes the same interface: it can send a message, receive one and generate a reply. What sits behind the reply is configurable. An AssistantAgent answers with an LLM and, by default, writes Python to solve problems. A UserProxyAgent answers by executing the code it receives, by asking a human, or both. When an agent receives a message, it replies automatically unless a termination condition is met, so two agents can carry a task forward without a central controller.

The second idea is conversation programming. Building an application means deciding two things: what each agent computes when it replies, and the control flow, meaning who speaks next and when the conversation stops. Both can be set in code (registered reply functions, termination checks, reply limits) or in natural language (a system message telling the assistant to reply "TERMINATE" when done). A GroupChatManager extends this to many agents by picking the next speaker, collecting its reply and broadcasting it to the rest.

The paper then builds six applications: math problem solving, retrieval-augmented question answering and code generation, ALFWorld household tasks, a supply-chain coding assistant called OptiGuide, a dynamic group chat, and conversational chess. Each shows a gain or a new capability. Read closely, the biggest improvements came from giving one agent a specific job, such as a safety check or a commonsense rule, that a single agent had been doing badly or not at all.

Customers ask for "multi-agent" systems because the term is popular. This paper is a good guide to when more agents help, because its own numbers are specific.

A second agent helps when it has a job the first was doing badly. OptiGuide's code writer and safety checker, run as two agents, caught unsafe code far better than one agent doing both, especially with the weaker model. The ALFWorld grounding agent added a rule the planner kept forgetting. A plain two-agent version of ReAct, by contrast, merely matched ReAct.

Termination is part of the design. Agents that reply to each other automatically will keep going. The paper's own pilot study found two-agent chats failing to terminate on up to 9 of 12 tasks. Every conversation needs a stop condition and a budget.

Start with a human in the loop. The authors' own guideline is to begin with human input always on, learn where the agents fail, and only then switch to autonomous mode.

Measure against a single, well-equipped agent. Several of AutoGen's comparisons pit it against systems with different models or no code execution. Your baseline should be one agent with the same tools and model.