Semantic Kernel vs Microsoft Agent Framework — Which for .NET in 2026?

The 2026 Dilemma: Semantic Kernel or Agent Framework?

Semantic Kernel and Microsoft Agent Framework are complementary layers of the Microsoft .NET AI stack, not competing alternatives. In 2026, Microsoft clarified the architectural hierarchy: Semantic Kernel provides the foundation layer (model connectors, plugin functions, vector memory, and streaming filters), while Microsoft Agent Framework (MAF 1.0 GA) provides the autonomous multi-agent layer (graph workflows, checkpointed agent state, and human-in-the-loop approvals).

Enterprise .NET architectures combine both frameworks to build scalable, resilient production AI systems.

Detailed Architectural Comparison

Code Example: Building Tools in Semantic Kernel

Semantic Kernel shines at defining clean, testable C# plugins with parameter schemas and dependency injection:

using System.ComponentModel;
using Microsoft.SemanticKernel;

public class InventoryPlugin
{
    [KernelFunction("check_stock")]
    [Description("Checks current warehouse inventory for a given SKU.")]
    public async Task<int> CheckStockAsync(
        [Description("The product stock keeping unit (SKU)")] string sku,
        CancellationToken ct = default)
    {
        // Simulated database inventory lookup
        await Task.Delay(10, ct);
        return sku.StartsWith("PROD-AI") ? 42 : 0;
    }
}

Code Example: Orchestrating Agents in Microsoft Agent Framework

Microsoft Agent Framework consumes the Semantic Kernel plugin above and coordinates multi-agent collaboration with built-in turn termination:

using Microsoft.Agents.AI;
using Microsoft.SemanticKernel;

// 1. Build underlying Semantic Kernel with Azure OpenAI / Microsoft Foundry
var kernel = Kernel.CreateBuilder()
    .AddAzureOpenAIChatCompletion("gpt-4o", "https://my-foundry.openai.azure.com/", "apiKey")
    .Build();

// 2. Register tools in Semantic Kernel
kernel.Plugins.AddFromType<InventoryPlugin>("Inventory");

// 3. Create autonomous MAF Agents sharing the kernel tools
var supportAgent = new ChatCompletionAgent
{
    Name = "CustomerSupport",
    Instructions = "Gather customer requirements and inquire about inventory when needed.",
    Kernel = kernel
};

var logisticsAgent = new ChatCompletionAgent
{
    Name = "LogisticsSpecialist",
    Instructions = "Verify stock levels using the Inventory tools and authorize shipments.",
    Kernel = kernel
};

// 4. Coordinate multi-agent group conversation
var groupChat = new AgentGroupChat(supportAgent, logisticsAgent)
{
    ExecutionSettings = new()
    {
        TerminationStrategy = new RegexTerminationStrategy(@"\b(SHIPMENT_APPROVED|OUT_OF_STOCK)\b")
    }
};

await groupChat.AddChatMessageAsync(new("Customer: I need 5 units of PROD-AI-10 immediately."));

await foreach (var message in groupChat.InvokeAsync())
{
    Console.WriteLine($"[{message.AuthorName}]: {message.Content}");
}

When to Use Which: Architectural Decision Matrix

Choose Semantic Kernel When:

1. You need a single intelligent assistant: Adding conversational chat or copilot sidebars to ASP.NET Core or Blazor apps.
2. You are building custom RAG pipelines: Ingestion, chunking, embedding generation, and vector retrieval with Azure AI Search, Qdrant, or Cosmos DB.
3. You want lightweight middleware: Logging, cost estimation, prompt caching, and token counting filters.

Choose Microsoft Agent Framework When:

1. You need autonomous multi-role teams: An analyst agent drafts a plan, a coder agent writes code, and a reviewer agent validates the output.
2. Your workflows span hours or days: Checkpointed workflows pause to wait for human manager approval and resume without losing context.
3. You require graph-based state transitions: Complex decision trees where agent outputs branch dynamically based on business rules.