Microsoft Agent Framework: Open-Source AI Agent SDK

If you have ever tried to get two AI agents to collaborate on a task reliably -- without one of them hallucinating a tool call, losing track of state mid-workflow, or requiring you to write 400 lines of retry logic -- you already understand the problem Microsoft Agent Framework was built to solve. MAF is not a wrapper or a chatbot kit. It is a production-grade orchestration runtime, and it arrived at General Availability in April 2026, right as the "agentic era" stopped being a marketing phrase and started being a real architectural challenge for engineering teams.

Microsoft announced the framework prominently at Open Source Summit North America 2026 and again at Build 2026, which opened on June 2 in San Francisco. The message both times was the same: agents are the new apps, and MAF is the open-source foundation Microsoft wants developers to build them on. Compared to what came before -- AutoGen's conversational loops, Semantic Kernel's plugin model -- MAF is a meaningful step up in architectural clarity and operational capability.

What Is the Microsoft Agent Framework and Where Did It Come From?

Microsoft Agent Framework -- MAF for short -- is an open-source SDK and runtime for building, orchestrating, and operating multi-agent AI systems. It lives on GitHub under the microsoft org, installs via pip install agent-framework for Python or NuGet for .NET, and is licensed permissively for both commercial and personal use.

MAF is best understood as the merger and evolution of two earlier Microsoft projects: AutoGen, which pioneered conversational multi-agent patterns in 2023, and Semantic Kernel, which provided a plugin-oriented integration layer for connecting LLMs to business tools. Both were useful research platforms. Neither was designed from the ground up for production deployments where durability, observability, and governance are non-negotiable. MAF is.

The framework reached v1.0 GA in April 2026 after roughly a year of preview. It carries forward the lessons of its predecessors but replaces their ad-hoc architectures with a consistent core: a graph-based workflow engine, a provider-agnostic agent runtime, and a middleware system that lets teams intercept, audit, and govern every agent action at every step of the pipeline.

What Are the Core Features That Make MAF Worth Using?

MAF 1.0 ships with a feature set that puts it well ahead of most open-source agent frameworks on production readiness. Here are the capabilities that matter most:

A companion project, the Agent Governance Toolkit, extends MAF with intent-based authorization and cost governance. Install it with pip install agent-governance-toolkit[full] and you get audit trails, per-agent spend caps, and safety rails that integrate directly with MAF's middleware pipeline -- a setup that matters a lot for regulated industries and enterprise deployments.

How Does MAF Compare to AutoGen, LangGraph, and Other Agent Frameworks?

The open-source agent framework landscape has gotten crowded fast. Here is where MAF sits relative to the most commonly compared alternatives:

• vs. AutoGen — AutoGen's strength was its conversational multi-agent model, where agents "talk" to each other to solve tasks. MAF retains that pattern as one option but adds the full graph-based workflow engine, durability primitives, and governance hooks that AutoGen lacked. Think of MAF as AutoGen grown up for production.
• vs. LangGraph — LangGraph is Python-only and tightly integrated with the LangChain ecosystem. MAF offers similar graph-based orchestration but adds .NET support, better native Azure integration, and a more explicit middleware contract. If your team uses .NET, MAF has no real competitor in the open-source space.
• vs. CrewAI — CrewAI targets a higher-level "roles and tasks" abstraction that is faster to prototype but harder to customize deeply. MAF sits lower in the stack, giving engineers precise control over every workflow node and every agent interaction -- at the cost of more initial setup.
• vs. Semantic Kernel — Semantic Kernel's plugin model is still useful for single-agent scenarios, but MAF supersedes it for multi-agent orchestration. Microsoft has signaled that MAF is the forward path for production agent systems built on their stack.

The key differentiator is production maturity. Most open-source frameworks are still research-grade tools that assume a happy path. MAF was designed with the assumption that agents will fail, providers will rate-limit, humans will need to intervene, and someone in compliance will eventually ask for an audit log. Check out our full open-source AI models index for more framework comparisons.

How Do You Install and Start Building with MAF?

Getting started with MAF takes about five minutes. The install path is standard for both runtimes:

From there, the official MAF docs on Microsoft Learn walk through building your first agent (single-turn), then adding a second agent (handoff pattern), then composing a full durable workflow with checkpointing. The sample library on GitHub covers 40+ scenarios including sequential, concurrent, and group collaboration workflows in both Python and C#.

For local development, MAF works out of the box with any OpenAI-compatible endpoint -- including Ollama running Llama 3, Mistral Small, or Qwen 3.6 on your own machine. You do not need an Azure account to develop and test. Azure comes in when you are ready to deploy to managed cloud infrastructure, at which point the Foundry Hosted Agents feature lets you push a production-ready agent with two additional lines of code on top of your existing local setup.

The framework also has an active Discord server (discord.gg/b5zjErwbQM) and a dedicated devblog at devblogs.microsoft.com/agent-framework where the engineering team posts deep dives on durable workflows, governance patterns, and upcoming features. If you are evaluating which framework to commit to for a multi-agent project in 2026, the ecosystem around MAF is worth factoring in alongside the technical capabilities. See our free AI tools directory for more open-source developer resources.

Is Microsoft Agent Framework Truly Free and Open Source?

Yes -- and this is one area where the story is unambiguously positive. The framework code is on GitHub, the Python package is on PyPI, and the .NET package is on NuGet. You can inspect, fork, contribute to, and deploy it in commercial products without paying Microsoft anything for the framework itself.

The costs that do apply are the ones you would expect from any AI project: the LLM API calls your agents make. If you route through Azure OpenAI, you pay Azure OpenAI pricing. If you use OpenAI directly, you pay OpenAI's API rates. If you self-host a model via Ollama or LocalAI, your only cost is the hardware. MAF itself adds zero overhead to that bill.

The optional Azure AI Foundry hosting (Foundry Hosted Agents) is a paid managed service -- but it is strictly optional. The framework runs fully on-premises, in your own cloud, or on a developer laptop. Microsoft's business model here is the classic open-core play: the SDK is free, and you pay if you want the managed cloud deployment layer. For most developers and small teams, the free local path is entirely sufficient for development and even production at moderate scale. Browse our free vs. paid AI comparisons for more context on open-source vs. commercial tradeoffs in 2026.

What Are MAF's Current Limitations and What Is on the Roadmap?

No framework ships perfect, and MAF is no exception. The most significant current limitation is the Agent-to-Agent (A2A) protocol, which is still in development as of v1.0. A2A will enable structured cross-runtime messaging between agents -- for example, a .NET orchestrator delegating to a Python specialist agent over a standardized protocol. Until A2A ships, cross-language agent collaboration requires manual API bridging, which is doable but adds friction.

The migration story from AutoGen is also still evolving. Microsoft published a migration guide covering the most common patterns, but teams with complex AutoGen workflows should budget time for the port. The APIs are conceptually familiar but not syntactically compatible -- you are not doing a find-and-replace, you are restructuring your agent topology around MAF's graph model.

On the roadmap, A2A 1.0 support is the highest-profile upcoming feature, alongside deeper Windows Local AI integration (local NPU execution without cloud round-trips), expanded MCP server catalog, and more built-in governance templates in the Agent Governance Toolkit. Microsoft has committed to a rapid release cadence through 2026 given Build's emphasis on agents as the central product theme for the year.