Microsoft Agent Framework: Microsoft's framework for AI agents

Two Microsoft projects merged into one

For several years Microsoft maintained two agent frameworks side by side: Semantic Kernel, aimed at enterprises, and AutoGen, born at Microsoft Research and focused on multi-agent experimentation. Microsoft Agent Framework brings them together. According to the official documentation, it combines AutoGen’s simple abstractions with Semantic Kernel’s enterprise features (session-based state, type safety, middleware, telemetry) and adds graph-based workflows to orchestrate multiple agents explicitly.

A first public preview shipped in October 2025, and version 1.0 for Python and .NET was released on April 2, 2026. A Go SDK lives in a separate repository and is in public preview.

What you can build with it

The framework is organized around four main areas:

• Agents: an AI agent backed by a model that calls tools and MCP servers.
• The harness agent: a batteries-included agent for long, multi-step tasks, with planning, to-do tracking, context compaction and memory.
• Workflows: graph-based flows (sequential, concurrent, handoff, group collaboration) with checkpointing, human approval and time travel back to an earlier state.
• Integrations: model providers, tools, evaluation, UI frameworks, and interoperability through the A2A protocol.

Getting started

Install the agent-framework package (Python) or the Microsoft.Agents.AI NuGet packages (.NET), pick a model client, then create an agent with a name and instructions: the official docs show a first agent in about ten lines. Microsoft’s DevUI then lets you test and debug your agents and workflows. If you are coming from Semantic Kernel or AutoGen, start with the matching migration guide.

Strengths and limits

Strengths: backing from a major vendor with a long-term support commitment, Python / .NET parity, explicit and durable workflows, built-in OpenTelemetry, and multi-provider model support.

Limits: documentation and samples lean toward the Azure ecosystem, which can be confusing outside Microsoft shops. New versions ship at a fast pace, so tutorials age quickly. Microsoft also reminds users that third-party models or servers are used at their own risk, and that checking where your data flows is your responsibility.

Alternatives

Outside the Microsoft ecosystem, LangGraph offers comparable graph-based orchestration, with failure recovery and human review.