Microsoft Agent Framework

Microsoft's open-source framework for building, orchestrating and deploying AI agents in .NET and Python.

FreeFree planOpen-source agent development framework (SDK)

No user reviews yet. Be the first to rate it.

Visit website
Microsoft Agent Framework website screenshot
Microsoft Agent Framework website

Quick Facts

DeveloperMicrosoft
Websitelearn.microsoft.com/en-us/agent-framework
Agent typeOpen-source agent development framework (SDK)
Best forDevelopers and enterprise engineering teams building production agents and multi-agent workflows in .NET or Python, including AutoGen and Semantic Kernel users.
Not ideal forBusiness users who want a no-code builder, or teams wanting a hosted product with a ready-made interface.
Pricing modelFree, open source (MIT license); pay for models and hosting
Starting priceFree
Pricing detailsThe framework is free under the MIT license. Costs come from the model providers and any Azure or other cloud services used to host agents.
Free planYes
ChannelsDeveloper SDK (.NET and Python packages)
What it does on its ownDevelopers define agents, tools and graph-based workflows with checkpointing and human-in-the-loop steps; autonomy depends on how each system is built.
IntegrationsMicrosoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, Ollama, GitHub Copilot, MCP, A2A, AG-UI
Platforms.NET, Python
Agent launched2025
Details last verified29 October 2026
CategoriesAgent Frameworks for Developers, Free AI Agents, Multi-Agent Platforms, Open-Source AI Agents
User ratingNo reviews yet

Microsoft Agent Framework reached version 1.0 in April 2026, giving .NET and Python developers stable APIs for building agents and multi-agent systems. It merges AutoGen, the multi-agent toolkit that came out of Microsoft Research, and Semantic Kernel, Microsoft’s enterprise SDK, into a single open-source project under the MIT license, with more than 14,000 GitHub stars.

For new agent work, Microsoft points developers here. AutoGen now gets only bug fixes and security patches, Semantic Kernel remains active, and both have migration guides. The framework is the code-first counterpart to Copilot Studio, Microsoft’s low-code agent builder, and suits teams that want to own every part of how their agents behave.

How It Works

Developers create an agent, give it type-safe function tools and connect it to a model. First-party connectors cover Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama. Middleware hooks, memory and context providers shape what the agent sees and does, which is where teams add logging, content filters or company-specific rules.

For larger jobs, graph-based workflows chain agents and functions with streaming, checkpointing and human-in-the-loop approval. Built-in orchestration patterns include sequential, concurrent, handoff, group chat and Magentic-One, a planner-led pattern from Microsoft Research. The framework supports the Model Context Protocol for tools, plus interoperability with the Agent2Agent protocol and AG-UI front ends. In preview are DevUI, a browser-based local debugger, along with hosted agents, memory, observability and evaluations in Microsoft Foundry, and integrations with the GitHub Copilot and Claude Code SDKs.

Pricing

The framework is free and open source. Teams pay for the models they call and for wherever they host agents, whether Azure, another cloud or their own servers. Packages install through pip for Python and NuGet for .NET, so there is no sign-up or license key.

What to Watch Out For

This is a developer toolkit, not a finished agent, so expect engineering work on hosting, security and monitoring. Some of the most useful operational features sit in preview and in Microsoft Foundry, which pulls teams toward Azure even though the core runs anywhere. Teams moving from AutoGen should plan migration time, since it is now in maintenance mode, and check that A2A support covers their needs before relying on it.

Recent News

Microsoft moved the framework to release candidate on February 19, 2026, and shipped version 1.0 for .NET and Python in early April 2026.

Features

  • Single-agent abstraction with type-safe function tools
  • Graph-based workflows with streaming and checkpointing
  • Orchestration patterns: sequential, concurrent, handoff, group chat and Magentic-One
  • Middleware hooks, memory and context providers
  • Connectors for major model providers
  • MCP and AG-UI support, with A2A interoperability
  • Migration guides from AutoGen and Semantic Kernel
  • DevUI local debugger (preview)

Pros and Cons

Pros

  • Free and MIT licensed
  • Stable 1.0 APIs for .NET and Python
  • Works with models from Microsoft, OpenAI, Anthropic, Google, AWS and Ollama
  • Clear successor path for AutoGen and Semantic Kernel users
  • Human-in-the-loop and checkpointing built into workflows

Cons

  • Requires programming skills
  • Several pieces, including DevUI and Foundry hosting, still in preview
  • Hosting, monitoring and model costs fall on you
  • Ties in most closely with Azure services
Write a Review