Microsoft’s agentic AI frameworks, Semantic Kernel and AutoGen are deeply collaborating to provide the best-in-class agentic developer experience. With Semantic Kernel’s enterprise ready AI capabilities, customers can already use and get support for building agent applications and, moving forward, we’ll align the multi-agent runtime in AutoGen (called autogen-core) with Semantic Kernel, allowing customers to create enterprise-ready multi-agent solutions.
Our two teams will be working quickly and diligently to get to this alignment, with the goal of making it available in early 2025. Meanwhile, for customers that want to ideate and experiment with multi-agents and use cutting edge agentic patterns, they can experiment with AutoGen. Customers should choose Semantic Kernel if they’re building agent production applications that need AI capabilities with enterprise-grade support.
Recently, the Semantic Kernel team introduced the Semantic Kernel Process Framework, allowing developers to embed AI into their workflows. Semantic Kernel Process Framework introduces the concepts of stateful, long-running process to agentic and human-in-the-loop process. Business process can be modelled and built in Semantic Kernel then distributed and scaled with Dapr and soon, Microsoft Orleans. Find out more about the Semantic Kernel Process Framework here: Integrating AI into Business Processes with the Process Framework.
The AutoGen team recently pre-released AutoGen 0.4, which also uses Microsoft Orleans and builds on learnings and insights acquired through the AutoGen community and across Microsoft’s work with agentic systems. This new version is a ground up re-design of an event-driven, distributed architecture that is x-language, composable, flexible, observable and scalable. Find out more about AutoGen 0.4 here: New AutoGen Architecture Preview | AutoGen. The AutoGen team also released Magentic-One a generalist multi-agent team that has shown state-of-the art performance on multiple benchmarks.
Both frameworks are converging into similar design principles for orchestrating multiple agents. Hence, the two teams will collaborate towards a unified multi-agent runtime, allowing customers experimenting with cutting edge agentic patterns and capabilities to have a seamless transition into an enterprise ready and supported environment.
FAQ to help you determine which framework aligns best with your needs:
What are the primary functions of Semantic Kernel and AutoGen?
- AutoGen: An open-source framework designed by Microsoft Research’s AI Frontiers Lab to build AI agent systems. It simplifies the creation and orchestration of event-driven, distributed agentic applications, enabling multiple LLMs and SLM’s, tools, and advanced multi-agent design patterns. AutoGen supports scenarios where multiple agents interact with each other to complete complex tasks autonomously or with human oversight. The event-driven and distributed architecture makes it suitable for workflows that require long-running autonomous agents that collaborate across information boundaries with variable degrees of human involvement. AutoGen currently supports C# and Python.
- Semantic Kernel: A production ready SDK that integrates large language models (LLMs) and data stores into applications, enabling the creation of product-scale GenAI solutions. Semantic Kernel supports multiple programming languages: C#, Python, and Java. Semantic Kernel has an Agent and Process Frameworks in preview, enabling customers to build single-agent and multi-agent solutions.
Which framework is more appropriate for production?
Semantic Kernel has reached version 1.0 across .NET, Python and Java, and is ready for production use. This milestone reflects a commitment to stability and non-breaking changes, providing developers with confidence in building enterprise-ready AI applications. Semantic Kernel is supported by Microsoft through services such as Microsoft Unified Customer Support.
Microsoft Research’s AI Frontiers Lab maintains AutoGen as an open-source framework, with a vibrant community that contributes and supports it. AutoGen is a vehicle for AI Frontiers to turn state of the art research into agentic capabilities and enable the development of AI applications that push today’s boundaries.
The AutoGen community is working towards a stable version of its multi-agent core runtime, which is the building block for all its agentic innovations. It is indeed this multi-agent core runtime that the two teams are trying to converge on, for a seamless transition between the two SDKs and more importantly to allow customers who start their journey in AutoGen to have a path for stability and Microsoft customer support into Semantic Kernel and Azure AI.
I am using AutoGen, what’s my path for product support?
If you’re using the multi-agent runtime from AutoGen, in early 2025 there will be an option to seamlessly transition into Semantic Kernel; this would alleviate the burden of having to productize it on your own and you can rely on an enterprise-ready runtime for your multi-agent solution.
If you’re willing to harden and make AutoGen enterprise ready with your overall solution, then you can continue using it with community support only. Customers might choose this path if they have a need for complex agentic patterns that are not yet available in Semantic Kernel or any other product.
Are there any resources to help me get started with these frameworks?
Yes, Microsoft provides resources for both frameworks:
Semantic Kernel:
- Microsoft Semantic Kernel GitHub Repository
- Semantic Kernel Blog
- Semantic Kernel Agent Framework on Microsoft Learn
AutoGen:
Leveraging the complementary capabilities of Semantic Kernel and AutoGen allows developers to explore innovative patterns while ensuring stability in production environments. We encourage you to take advantage of these robust frameworks to realize your AI development goals. Stay tuned for more about how we’re building a pipeline to you bring your AutoGen agentic applications into Semantic Kernel! Thank you for being part of our community, and together, let’s drive the future of AI!
Congratulations! This sounds like the best strategy and way forward 🙂
With this I understand that the Semantic Kernel process framework will be adopted as the multiagent core, yes?
With this I understand that the move towards v0.4 on AutoGen is to be ready for this “Assemble”, is that correct?
And last, on the early 2025, that sounds like Build 2025 right? 😉