Microsoft Guides C# Developers in Shifting from Semantic Kernel to Agent Framework

Microsoft’s new Agent Framework marks a decisive shift from the dual‑SDK model that has defined .NET AI development for the past two years. At .NET Conf 2025 the company announced that the framework has reached Release Candidate status, merging the previously separate Semantic Kernel and AutoGen libraries into a single, type‑safe SDK for C# agents (Brian Spann, March 2). The consolidation eliminates the “confusing overlap” that developers complained about, according to Spann, and provides a stable API surface that Microsoft expects to ship as GA in Q1 2026. By treating the agent as a first‑class object—complete with name, instructions, and built‑in capabilities—the new model reduces boilerplate and aligns with Microsoft’s broader “agent‑centric” strategy outlined in recent VentureBeat coverage of the retirement of AutoGen (Emilia David, VentureBeat).

The most visible change for developers is the simplification of agent creation. In the legacy Semantic Kernel approach, a developer had to construct a Kernel builder, attach an Azure OpenAI chat completion service, and manually manage a ChatHistory object before invoking GetChatMessageContentsAsync (Spann). The Agent Framework replaces that multi‑step pipeline with a single fluent call: an OpenAI client is instantiated with Azure credentials, then wrapped in AsAIAgent to produce an agent instance that can be invoked directly via RunAsync (Spann). This shift not only cuts the amount of code required but also embeds human‑in‑the‑loop and streaming support out of the box, features that were previously optional extensions in AutoGen (David). The new SDK’s “Capability Agent Framework” label signals that these capabilities are now part of the core, rather than add‑ons.

Function tools—Microsoft’s term for plugins that expose external capabilities—receive a more structured migration path. Under Semantic Kernel, plugins were defined as plain C# classes annotated with [KernelFunction] and [Description] attributes, then registered with the kernel (Spann). The Agent Framework retains the concept of function tools but enforces type‑safe definitions and integrates them directly into the agent’s contract, eliminating the need for a separate kernel registration step. This change is highlighted in Spann’s side‑by‑side code comparison, where the WeatherPlugin example would now be declared using the framework’s ToolDefinition API, ensuring compile‑time validation of input and output schemas. VentureBeat’s analysis of “agentic design patterns” notes that such strict typing is essential for enterprise reliability, especially when agents orchestrate multi‑provider calls across Azure OpenAI, Anthropic, Ollama, and other services (Ben Dickson, VentureBeat).

Beyond syntax, the Agent Framework introduces a unified runtime for multi‑agent workflows. AutoGen previously offered “group chat” and “handoff” scenarios, while Semantic Kernel focused on orchestration via plugins. The new SDK consolidates these patterns under a single workflow engine that supports sequential, concurrent, and group‑chat topologies without additional libraries (Spann). Developers can now compose complex interactions—such as a sales‑assistant agent delegating pricing queries to a pricing‑engine agent—using the same fluent API that defines a single agent. This interoperability is reinforced by the framework’s adherence to emerging standards like A2A, MCP, and AG‑UI, which Microsoft cites as “interoperability” guarantees (Spann).

The migration guidance itself is pragmatic. Spann advises developers to start with a “pilot conversion” of low‑risk agents, leveraging the RC’s stable API to validate behavior before the full GA rollout. Because the Agent Framework abstracts credential handling through Azure Identity, existing Azure OpenAI deployments can be re‑used without code changes, easing the transition for enterprises that have already invested in Microsoft’s cloud AI stack. The article also notes that the framework’s built‑in observability hooks—exposed via the Microsoft .Diagnostics namespace—provide richer telemetry than the ad‑hoc logging patterns common in Semantic Kernel projects (Spann). This aligns with VentureBeat’s coverage of the broader push toward “governed enterprise AI agents,” where visibility and compliance are becoming mandatory requirements (David).