Semantic Kernel and AutoGen Become the Microsoft Agent Framework
Microsoft Agent Framework entered public preview on October 1, 2025, converging Semantic Kernel and AutoGen into one SDK. Both predecessors are now in maintenance mode. We summarize the verified facts, map the .NET programming model changes, and outline a migration sequence that avoids maintenance-mode debt before the framework reaches general availability in 2026.
The Problem With Betting on Two Frameworks
Since 2023, .NET teams building LLM applications on the Microsoft stack have faced a choice between two SDKs. Semantic Kernel delivered enterprise plumbing: dependency injection, OpenTelemetry, stable APIs. Its agent framework reached general availability in April 2025 with Semantic Kernel 1.45 for .NET. AutoGen, a Microsoft Research project, delivered multi-agent orchestration; version 0.4 shipped in January 2025. The two overlapped in scope but shared neither abstractions nor a roadmap.
Every architecture decision was therefore a bet on which SDK Microsoft would carry forward. On October 1, 2025, Microsoft answered: neither. Both converge into the Microsoft Agent Framework. That answer removes the ambiguity and creates a concrete engineering task: planning the transition before maintenance mode turns into technical debt.
What Microsoft Announced in October 2025
Microsoft Agent Framework entered public preview on October 1, 2025. It is an open-source SDK and runtime for building and orchestrating AI agents, developed on GitHub under microsoft/agent-framework, with full support for .NET and Python. It combines AutoGen's multi-agent orchestration with Semantic Kernel's enterprise foundations and builds on Microsoft.Extensions.AI as its abstraction layer. It supports the Model Context Protocol, the Agent2Agent protocol, OpenAPI-first tool integration, and OpenTelemetry-based observability. Pluggable memory modules cover Redis, Qdrant, Elasticsearch, Postgres, and custom stores.
The .NET packages ship as Microsoft.Agents.AI plus companion packages such as Microsoft.Agents.AI.Workflows, all versioned 1.0.0-preview; the release current at the time of writing is 1.0.0-preview.251125.1 from November 26, 2025. Microsoft describes the framework as Semantic Kernel v2.0, built by the same team. That framing is accurate: this is a major-version rewrite, not a new product category.
Maintenance Mode Is Not End of Life
Microsoft's statements are specific. Semantic Kernel and AutoGen remain in maintenance mode: no new feature investment, but continued bug fixes, security patches, and stability updates. No breaking changes are planned. Semantic Kernel 1.x will be supported for at least one year after Agent Framework reaches general availability, and selected existing features will still be taken to GA. Migration guides for .NET and Python were published alongside the announcement.
Maintenance mode is therefore not an emergency. It is a slow leak. New model capabilities, new protocols, and new integrations will land only in Agent Framework. A Semantic Kernel codebase does not break in 2026; it falls behind. The cost is measured in features you cannot adopt, not in builds that fail.
The New Programming Model in .NET
The .NET surface is smaller and flatter than Semantic Kernel's. The base abstraction is AIAgent. The workhorse is ChatClientAgent, which wraps any IChatClient from Microsoft.Extensions.AI. One agent type replaces ChatCompletionAgent, OpenAIAssistantAgent, and AzureAIAgent. The Kernel object is gone; agents are constructed directly from a chat client and registered in dependency injection like any other service.
Provider breadth now comes from the IChatClient ecosystem rather than from first-party connectors: the AWS SDK, Anthropic.SDK, and ONNX Runtime GenAI each ship their own IChatClient implementation. Existing Semantic Kernel IChatCompletionService instances adapt via the AsChatClient extension. Orchestration moves into graph-based workflows in Microsoft.Agents.AI.Workflows, with checkpointing and human-in-the-loop steps built in.
Mapping Semantic Kernel to Agent Framework
Microsoft publishes an official migration guide for .NET covering packages, dependency injection, options, and thread management. Semantic Kernel's KernelFunction is already an AIFunction, so tool code largely carries over. In our experience the chat-agent layer migrates mechanically; group-chat and custom orchestration logic requires redesign as workflow graphs. The core concepts map as follows in the table below.
Expect the mechanical part to take days, not months, for a typical service with a handful of agents. Expect the workflow redesign to dominate the budget when AgentGroupChat or custom termination strategies are involved. Prompt templates and planner-era code have no one-to-one equivalent and should be rewritten as tools or workflow steps.
| Semantic Kernel | Microsoft Agent Framework |
|---|---|
| Kernel instance | Not required |
| ChatCompletionAgent / OpenAIAssistantAgent / AzureAIAgent | ChatClientAgent |
| IChatCompletionService | IChatClient (Microsoft.Extensions.AI) |
| KernelFunction plugins | AIFunction tools |
| AgentGroupChat orchestration | Graph-based workflows |
| Microsoft.SemanticKernel packages | Microsoft.Agents.AI packages |
What We Recommend Before GA
Our recommendation follows Microsoft's own guidance from October 7, 2025, which is unusually candid. Existing Semantic Kernel projects that must ship now should stay on Semantic Kernel. New projects that can ship after GA should start on the Agent Framework preview. Nobody should port a stable production system while the API is still moving between preview releases.
The lowest-risk step works today: move your code onto Microsoft.Extensions.AI abstractions. Semantic Kernel has supported IChatClient natively since mid-2025, and Agent Framework builds on the same types. Wrap agent construction behind your own interfaces, pin package versions, and inventory every direct use of Kernel, KernelFunction, and connector-specific types. That inventory is your migration estimate.
Be clear about what the migration does not buy you. It changes the SDK surface, not model behavior. Evaluations, guardrails, prompt quality, and cost control carry over unchanged — as does their maintenance. And during preview, .NET and Python are not guaranteed feature parity; verify that the features you need exist in the .NET packages, not only in the announcements.
Outlook From December 2025
As of today, Microsoft has committed to several months of preview but no GA date. Ignite on November 18, 2025 showed the direction: Hosted Agents in Foundry Agent Service in public preview, durable execution, Purview integration, and support for the AG-UI and ChatKit front-end protocols. We expect general availability in the first half of 2026, which would run the Semantic Kernel support commitment into 2027 at minimum.
We also expect the usual preview arc: some announced features will slip past GA, and early adopters will absorb breaking changes between preview releases. That is the price of the head start. Our prediction: by the end of 2026, Agent Framework is the default answer for agents on .NET the way ASP.NET Core is for web — and teams that spent 2026 on the Extensions.AI groundwork will migrate in weeks, while teams that waited will migrate under deadline.
Sources
- Introducing Microsoft Agent Framework — Microsoft Azure Blog (1 Oct 2025)
- Introducing Microsoft Agent Framework (Preview) — .NET Blog (1 Oct 2025)
- Semantic Kernel and Microsoft Agent Framework — Shawn Henry (7 Oct 2025)
- Microsoft retires AutoGen and debuts Agent Framework — VentureBeat (1 Oct 2025)
- Semantic Kernel Agents are now Generally Available — Microsoft (4 Apr 2025)
- Build and scale AI agents with Microsoft Foundry — Microsoft Azure Blog (18 Nov 2025)