One of the most popular open-source frameworks of 2023, which explored the application of AI agents and Multi-agent collaboration, has recently seen team separation and the formation of new products.
Back in September, a few original creators of Autogen parted their way with Microsoft creating a GH fork of the original project. Eventually, in November they settled on creating a new AG2 GitHub Org and new repo (this time a fresh one not leading back to the original repo) while inheriting PyPi autogen and pyautogen packages (that's where they push their updates) and Discord channel.
At the same time, Microsoft has introduced a complete rewrite called Autogen 0.4 which is available at the master branch of the original repo. Yet there's a legacy 0.2 version maintained by the community and available in a separate branch of the original repo. Both versions are available in a new PiPy repo called autogen-agentchat.
Recently Microsoft has announced that they are planning on brining Autogen's 0.4 multi-agent runtime into Semantic Kernel (promising seamless transition from Autogen to SK), essentially introducing the 4th reincarnation of Autogen. Starting in early 2025 they intend to ship agent runtime as part of SK repo providing production support while leaving the maintenance of the original repo to the community.
That's quite a mouthful, here's my attempt to order the links:
1. Complete rewrite, Autogen 0.4 with different architecture
- Developed in the master branch on the original Microsoft repo: https://github.com/microsoft/autogen
- Not backward compatible
- Maintained by community (? and MS for the time being)
- Available as autogen-agentchat==0.4.0.dev6 PyPi package (marked as Pre-Release version) and autogen-core
2. Legacy Autogen 0.2
- Developed in 0.2 branch in the original Microsoft repo https://github.com/microsoft/autogen/tree/0.2
- Backward compatible
- Maintained by community
- Available as autogen-agentchat~=0.2 PyPi package (marked as Release version, currently at version 0.2.38)
3. AG2 branched out by original creators:
- New repo: https://github.com/ag2ai/ag2
- Community Discord: https://discord.gg/pAbnFJrkgZ
- Backward compatible
- Available as
ag2,autogen, andpyautogenPiPy packages (pointing to identical code) - Currently at v0.3.2 and no breaking changes are planned for 0.4.x
4. Semantic Kernel that will eventually merge parts of Autogen 0.4
- Repo: https://github.com/microsoft/semantic-kernel
- Maintained by Microsoft
- Based on how I read the SK blog post, this is a recomended option in 2025, offering seamless migration from Autogen 0.4, better support and stability
P.S>
Here's a statement from MS Autogen camp regarding "clones". Seems they wanted to have pyautogen PyPi package pointing to their repo ("The current pyautogen package isn’t affiliated with Microsoft AutoGen, and admin access is blocked for us"), as well as Discord.
After all, there're 5 packages related to Autogen:
- AG2:
autogen,pyautogen,ag2, - MS:
autogen-agentchatandautogen-core
Top comments (10)
Hi @maximsaplin - I'm a MSFT employee, one of the architects of autogen 0.4, the article is a little inaccurate in a few ways, particularly the mention of "community maintenance". AutoGen is very much alive, very much still being invested in and maintained by a growing team at MSFT from across multiple divisions, is underlying agent systems in multiple 1p products including many announced yesterday, and yes will be converging with Semantic Kernel to offer a unified, scalable, distributed, xlang, standards based agent runtime. We are continuing to maintain the 0.2 branch but new features will show up in main. The differentiation for SK is around support, not stability, scale, etc.
Thanks for clarification! Quick question, would the below be an accurate update?
...
Maintained by Community and Microsoft (moving to community only support in early 2025)
...
Legacy Autogen 0.2
Maintained by Community
Not exactly.
AutoGen 0.4 - new event-driven actor model architecture, designed for scale, resilience, flexibility, forward looking feature development, driven by customer and community feedback, support x-lang agent communication, will be supported in Semantic Kernel as well. Built by dedicated teams at MSFT, continuing and growing investment, used in multiple MSFT products and services (not moving to maintained by community - not sure where that's coming from)
Semantic Kernel - a general purpose AI application SDK that will take a dependency on the Autogen runtime for multi agent. You can seek enterprise support for SK.
AutoGen 0.2 - the previous autogen arch, constrained to chat scenarios, synchronous, not scalable. still maintained by msft, but no new feature dev.
AG2 - unrelated to MSFT fork of autogen 0.2 made by a former employee - no significant new features or company behind it.
In your original comment you touched the question of support, Community vs Microsoft - that's where I can see possible inaccuracy. Any other parts? The rest of your comments seem to be expanding of what is already in the post, focusing on the MSFT side.
To be clear, I am not taking sides. There're currently 2 almost identical products: AG2 and AutoGen 0.2. And there's a lot of confusion due to uncoordinated and challenging split. There's no purpose (in my post) to make comparisons and tell how various derivatives/branches/variants of AutoGen are different - those interested can follow the links.
I just want to make it clear that Microsoft is going to continue to invest in (and grow) its investment in autogen 0.4, and is still involved in maintaining Autogen 0.2 - "community maintainance" implies it being abandoned, which is not the case. The distinction with SK is around Commercial Support, which is a feature available with SK.
and yes, uncoordinated and challenging are accurate in that the former employee is unilaterally implying that they are the official autogen project from Microsoft when it could have been easy to signal that it was indeed a distinct project.
My experience from last year/early this year (when I contributed to Autogen) was such that community (including independent contributors such as myself) drove a significant portion of changes, if not the major part. Folks from MS (as far as I remember) did extra mile.
Autogen might be one of a few OS projects that did see a lot of development from contributors not employed/assigned by companies sponsoring the development.
"Maintained by community" seemed to work fine (with the exception of, my personal opinion, code base becoming disorganised and not developer friendly). I suppose the momentum is still there and community support doesn't mean abandoned...
I see - so we were just each getting different meaning out of those words.
In the end the thing I wanted to land is that absent the confusion created by the announcement the project is going from strength to strength - there are plenty of msft teams that are really invested in it now and that guarantees a certain amount of ongoing growth and evolution, which we are excited about. The way in which the "AG2" project was announced seems intended to create doubt and FUD and that's too bad because the opportunity to contribute and work together has always been there, as you have observed. Looking forward to working with you on it further!
Well-structured and detailed explanation of the current state of Autogen and its evolution into multiple iterations and projects. The breakdown of the four key versions and their respective links is especially helpful in navigating a complex topic. Well done!
Also for those interested in this topic you may find this useful too: sdh.global/blog/ai-ml/microsoft-au...