AutoGen

What AutoGen is

AutoGen is a framework for building AI agents and applications ^[1]. An agent here means a program built around a large language model (LLM) — a model that generates text — that is given a role, instructions and some way to act, rather than a single prompt that produces a single answer.

The framework's paper describes agents that are customizable and conversable, and that can combine LLMs, human input and tools ^[2]. "Conversable" is the central idea: an agent is something you can send messages to, and something that can send messages back, so several agents can be placed in the same conversation and take turns. Tools and human input can enter that conversation as well, which means an agent is not limited to talking — it can call something that does work, or wait for a person to answer.

Where it came from

The paper, titled "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation," was submitted to arXiv on 16 August 2023 ^[2]. Its authors include Qingyun Wu, Gagan Bansal, Doug Burger and Chi Wang ^[2].

One of the paper's stated points is that conversation patterns can be programmed in both natural language and computer code ^[2]. In practice this means the description of how agents should behave — who speaks when, when to stop, when to hand over to a human — can be written as ordinary prose that the model reads, or as program logic that the runtime executes, and the two can be mixed in one system.

How it is organized

The current framework is split into two parts ^[1]. AgentChat is a programming framework for conversational single-agent and multi-agent applications; using it requires Python 3.10 or newer ^[1]. Core is an event-driven framework aimed at scalable multi-agent systems, a different starting point for people who need control over how messages are routed and processed rather than a ready-made chat loop ^[1].

Two further pieces sit on top. AutoGen Studio is a web-based interface for prototyping with agents without writing code, built on AgentChat ^[1]. Extensions connect the framework to outside services, and the documented examples include an McpWorkbench for MCP servers and a Docker command line code executor ^[1] — the latter a way for an agent to run code through a command line inside a Docker container, so that generated code executes in an isolated environment rather than directly on the host machine.

Why the conversation idea matters

Most early use of large language models was single-turn: one instruction in, one answer out. Splitting a task across several agents changes that. A common arrangement is to give one agent the job of producing something and another the job of checking it, or to separate planning from execution, letting the agents exchange messages until the result is good enough. The framework supplies the plumbing — message passing, turn-taking, the point at which a human can step in — so that the pattern itself can be described instead of hand-coded from scratch each time.

AutoGen's place in the field follows from that: it is one of the frameworks that made multi-agent conversation a standard way to structure LLM applications, alongside a code-first library, a no-code prototyping interface and connectors to external tools and services ^[1].