Microsoft Agent Framework 1.0 GA Released — What Changed for .NET Developers

The Unified Agent Runtime for .NET

Microsoft Agent Framework 1.0 has officially reached General Availability (GA), establishing the enterprise standard for autonomous, multi-agent AI systems in .NET. Following months of release candidates, Microsoft unified the multi-agent conversational patterns from AutoGen and the type-safe orchestration from Semantic Kernel into a single, production-grade library under the Microsoft.Agents.* namespace.

For enterprise C# developers, this milestone ends architectural fragmentation. Whether you are building sequential reasoning chains, hierarchical supervisor agents, or long-running workflows with human approvals, MAF 1.0 provides the officially supported framework.

What’s New in the 1.0 GA Release?

1. Unified Package Architecture

Experimental packages (Microsoft.SemanticKernel.Agents.*) and AutoGen .NET wrappers have been consolidated into clean, modular NuGet packages:

dotnet add package Microsoft.Agents.AI --version 1.0.0
dotnet add package Microsoft.Agents.Core --version 1.0.0
dotnet add package Microsoft.Agents.Storage.Cosmos --version 1.0.0

2. Checkpointed State & Workflow Resumption

In enterprise environments, agent tasks can run for hours or wait days for a human manager to approve a financial transaction or code pull request. MAF 1.0 introduces native checkpointing:

using Microsoft.Agents.AI;
using Microsoft.Agents.Core;

// Configure checkpoint storage backed by Azure Cosmos DB or Redis
var stateStore = new CosmosCheckpointStore("AccountEndpoint=...;AccountKey=...");

var agent = new ChatCompletionAgent
{
    Name = "CreditUnderwriter",
    Kernel = kernel,
    StateStore = stateStore
};

// Checkpoint saved automatically after each reasoning step
var session = await agent.ResumeSessionAsync("loan-app-9842");

3. Declarative Workflow Graphs

Beyond basic chat groups, MAF 1.0 introduces WorkflowGraph. Developers can construct directed acyclic graphs (DAGs) and cyclic state machines where agents pass typed payloads:

var workflow = new WorkflowGraphBuilder()
    .AddStep("ExtractRequirements", requirementAgent)
    .AddStep("GenerateCode", developerAgent)
    .AddStep("SecurityScan", securityAgent)
    .AddConditionalEdge("SecurityScan", 
        condition: result => result.HasVulnerabilities, 
        targetStep: "GenerateCode")
    .Build();

4. Built-in OpenTelemetry & AI Tracing

MAF 1.0 emits standard W3C trace contexts and semantic AI conventions. Every agent turn, LLM prompt, tool call duration, and token usage count is surfaced in Azure Application Insights, Aspire Dashboard, or Datadog without third-party shims.

Migration from Semantic Kernel Agents

If you have existing code using the experimental Microsoft.SemanticKernel.Agents preview packages, upgrading to MAF 1.0 GA requires three simple adjustments:

Getting Started

To explore hands-on implementation patterns with MAF 1.0, read our comprehensive guides:

• Microsoft Agent Framework 1.0 for .NET — Complete C# Guide
• Semantic Kernel vs Microsoft Agent Framework — Which for .NET in 2026?
• Migrate Semantic Kernel to Microsoft Agent Framework 1.0 — C# Step by Step