AutoGen’s successor is reorganizing the furniture before v1.0

What it does

AG2 is a Python framework for wiring up multiple LLM-powered agents so they can message each other, call tools, and execute code to complete tasks. It grew out of AutoGen and is now maintained by a dynamic group of volunteers from several organizations. The core abstraction is the ConversableAgent, which acts as a base class for every participant in a workflow, whether it is an AI model, a tool executor, or a human proxy.

The interesting bit

The framework treats multi-agent coordination as a conversation problem rather than a rigid pipeline: you can set up swarms, group chats, nested chats, or sequential patterns, and let an AutoPattern manager decide who speaks next. A built-in human-in-the-loop UserProxyAgent can halt the chatter when a real person needs to approve or redirect the output.

Key highlights

• Built-in conversation patterns: swarms, group chats, nested chats, and sequential chats.
• ConversableAgent base class unifies LLM agents, tool runners, and human proxies under one messaging interface.
• Supports tool registration, code execution, structured outputs, and RAG.
• Human-in-the-loop via UserProxyAgent with configurable termination conditions (e.g., checking for “DONE!” or “APPROVED”).
• The existing framework is being tidied up and moved to maintenance mode while the autogen.beta framework becomes the official v1.0 release.

Caveats

• The project is in a transitional phase; the current API is being deprecated and replaced by the beta framework before v1.0 lands.

Verdict

Worth a look if you are prototyping multi-agent LLM applications and want a conversation-centric orchestration layer with human oversight. Skip it if you need a stable, battle-tested production framework right now, because the ground is shifting beneath your feet.