What does the new Microsoft Agent Framework mean for Semantic Kernel? Support status & migration guidance

[CynthiaLuijkx]

With Microsoft announcing the new Agent Framework this week, I’m trying to understand what this means for the future of Semantic Kernel (SK) and for teams currently building on SK.

Some questions:

1. Is SK still under active development as a standalone SDK, or should we consider Agent Framework the long-term home for agentic scenarios?
2. What level of support/maintenance should we expect for SK (bug fixes, security patches, minor features) over the next 6–12 months?
3. For teams that can migrate, do you recommend moving to Agent Framework now, or waiting for a specific milestone (e.g., GA)?

[wim-gdwi]

One of the major issues I see for the time being is one of the major advantages we had with the use of SK: the support for various vendors via the connectors. Even though it is not completely abstracted away in SK, it is possible. For the new agents framework, I see no such libraries. I suspect the goal is that everything should go via the Azure Foundry "way of working" where you can also deploy various models, but, that is not the same imho. Going directly to a provider has certain advantages (and of course disadvantages) but we had a choice.

I see in the agents documentation and announcements that there should be more abstractions and that you do not have to differentiate anymore between OpenAi..., Mistral, ... classes and other things alike. So that should make it a lot easier. But honestly, that is of course also a remark that is true if the only thing supported is Azure AI Foundry and OpenAI, and you have to deploy the other models in Azure AI Foundry. If that would be the case, then the new agents library is a .NET/Python library for Azure AI foundry more than a general use abstraction layer that SK was.

Probably my focus is on this part because we actually use these different providers directly in our software and it works quite well. If it turns out that the different connectors are no longer an option in the new agents framework, I'm afraid I'll have to look for another .NET solution to support these different vendors, worst case, implement each of them ourselves directly.

So our software is really depending now on how SK is evolving and being supported... unfortunately already noticing the effort in it is minimized and PRs remain a long time open (if ever integrated).

[stephentoub]

For the new agents framework, I see no such libraries.

There are many, they just show up differently. ChatClientAgent works with any IChatClient (the evolution of SK's IChatCompletionService), and the model with IChatClient is not that all implementations are maintained by a single entity but rather that every relevant nuget package carries its own implementation. Want to use AWS Bedrock Runtime? The AWS SDK includes an IChatClient. Want to use Anthropic? The Anthropic.SDK nuget package includes an IChatClient. Want to use ONNX Runtime GenAI? That nuget package includes an IChatClient. Etc. (Also, every existing SK IChatCompletionService can be used as an IChatClient via the AsChatClient extension method built into SK.) You can see an example of that here:
https://github.com/microsoft/agent-framework/blob/62854197aa30c744562a9ca2d8585bc8583e9118/dotnet/samples/GettingStarted/Workflows/_Foundational/05_MultiModelService/Program.cs#L12-L57
where three different services are used for three different agents, none of them with Foundry.

[wim-gdwi]

Thx for your feedback and remark. Didn't see that immediately but will definitely investigate this!