AG2

AG2 is the community fork of Microsoft's AutoGen. When Microsoft shifted focus to its own Agent Framework in 2025, the AutoGen community took the code and kept building. The result: AG2, an open-source multi-agent framework where AI agents talk to each other in structured conversations to solve problems.

The core idea is simple. You define agents with different roles (coder, reviewer, planner) and let them converse. A coder agent writes Python. A reviewer agent critiques it. A human jumps in when needed. The framework manages message passing, turn-taking, and termination conditions. It's like a group chat, but the participants are LLMs.

AG2 vs. the Competition

CrewAI takes a task-oriented approach: define tasks, assign agents, execute sequentially. AG2 takes a conversation-oriented approach: agents negotiate, iterate, and course-correct through dialogue. For code generation and research workflows where agents need to go back and forth, AG2's model works better. For linear task pipelines, CrewAI is simpler. LangGraph gives you the most control. It uses explicit state machines with nodes and edges. More powerful, more complex. If you need precise control over agent routing and state transitions, LangGraph wins. If you want agents that "just talk" and figure things out, AG2 is more natural.

The Fork Situation

Here's the awkward part. Microsoft still maintains AutoGen (now part of their Agent Framework). AG2 is the community fork. Microsoft has publicly described AG2 as "just one of thousands of forks." This creates ecosystem fragmentation. Libraries, tutorials, and StackOverflow answers may reference AutoGen, AG2, or both. Check import paths carefully.

GroupChat: The Killer Feature

AG2's GroupChat puts multiple agents in a shared conversation with a selector that determines who speaks next. You can run 5, 10, or 20 agents debating solutions. A "Captain Agent" can spawn sub-agents on the fly. This pattern works well for complex research tasks where you don't know the right approach upfront.

Pricing

• Open Source: $0. MIT-adjacent licensing. Full framework access. Self-hosted.
• Enterprise Support: Contact Microsoft (though note: Microsoft may direct you to their Agent Framework instead)

Your costs are LLM API calls. A multi-agent conversation with 5 turns between 3 agents using GPT-4 costs roughly $0.30-1.00 depending on context length. That adds up fast in production.

Source: ag2.ai

The Pricing Gotcha

AG2 itself is free, but conversation-driven coordination is token-hungry. Each agent response costs API tokens. A 10-turn conversation between 3 agents generates 30 LLM calls. Compare that to single-agent approaches that solve the same problem in 1-3 calls. Budget for 3-10x the token cost of a single-agent solution.

Common Questions

Should I use AG2 or AutoGen? If you want community-driven development independent of Microsoft's roadmap, use AG2. If you want Microsoft enterprise support and Semantic Kernel integration, stick with AutoGen/Agent Framework. Is AG2 production-ready? For internal tools and prototypes, yes. For customer-facing production systems, proceed with caution. There's no built-in observability platform. You'll need to add your own logging, tracing, and monitoring. Can I migrate from AutoGen to AG2? Yes. AG2 maintains backward compatibility. Most AutoGen code runs on AG2 with minimal changes.

What Real Users Say

Developers praise AG2 for being "simpler, cleaner, and more flexible" than alternatives for rapid iteration. The conversation-driven approach gets called "the right level of abstraction" that "gets out of the way and lets you hack." Complaints center on the fork confusion, lack of production tooling, and higher token consumption compared to LangGraph or CrewAI.

Value Math

AG2 costs $0 in licensing. Your spend is API tokens. A typical multi-agent workflow costs $0.50-2.00 per run with GPT-4. Running 100 tasks/day = $50-200/month in API costs. CrewAI has a similar cost profile but adds a $200/month Enterprise tier for managed hosting. LangGraph Cloud starts at $0 for 1M nodes but charges for compute beyond that. AG2 keeps you in control of infrastructure and costs.