Microsoft Agent Framework

What is Microsoft Agent Framework?

Microsoft Agent Framework (MAF) is Microsoft’s open-source SDK for building and orchestrating AI agents and multi-agent workflows, unifying two previously separate Microsoft stacks — AutoGen’s multi-agent orchestration and Semantic Kernel’s enterprise plumbing — into one supported framework for .NET, Python and Go. It reached a production-ready 1.0 GA release on 2026-04-03, and Microsoft has placed both AutoGen and Semantic Kernel into maintenance mode (bug and security fixes only) specifically to push new development toward it. Released under the MIT licence on GitHub (microsoft/agent-framework, 12.4k stars and counting), it ships consistent APIs across languages, native support for the Model Context Protocol (MCP) and the Agent2Agent (A2A) protocol, and an optional managed hosting layer (Foundry Hosted Agents) alongside the free, fully self-hostable core.

What does Microsoft Agent Framework do?

You build an agent from a chat-client-backed model (Azure OpenAI, OpenAI, Microsoft Foundry, Anthropic, Ollama or a custom provider), give it instructions and tools, and compose multiple agents into a workflow using graph-based orchestration patterns — sequential, concurrent, handoff or group-chat-style collaboration. A production-focused "Agent Harness" layer adds shell and filesystem access, human-in-the-loop approval flows, automatic context/token compaction and built-in providers (FileMemory, FileAccess, Todo, Shell) so an agent can act on a real environment safely rather than only returning text. A newer "CodeAct" mode has the model write short Python programs that execute tool calls inside isolated Hyperlight micro-VMs instead of making tool calls one at a time — Microsoft’s own BUILD 2026 benchmarks report 52.4% faster completion and 63.9% fewer tokens on multi-step workloads. You can self-host the whole stack for free, or deploy to Foundry Hosted Agents for managed, scale-to-zero infrastructure with per-session isolation and OpenTelemetry/Application Insights observability built in. First-party migration guides walk existing AutoGen and Semantic Kernel projects through the move, and a GitHub Copilot SDK integration lets an agent reuse Copilot’s own shell/file tools directly.

How Microsoft Agent Framework works

1. 1Create an agent backed by a chat-client provider — Azure OpenAI, OpenAI, Microsoft Foundry, Anthropic, Ollama or a custom implementation of the AIAgent/BaseAgent interface.
2. 2Give it instructions and tools (ordinary functions), and optionally layer on the Agent Harness for shell/filesystem access, human-in-the-loop approval and automatic context compaction.
3. 3Compose multiple agents into a workflow using a graph-based orchestration pattern — sequential, concurrent, handoff or group collaboration — or connect to remote agents via the A2A protocol.
4. 4Self-host the workflow yourself, or deploy it to Foundry Hosted Agents for managed, scale-to-zero infrastructure with built-in OpenTelemetry observability.

Who is Microsoft Agent Framework for?

• .NET or Python teams already running AutoGen or Semantic Kernel who need a supported migration path onto Microsoft’s live successor
• Developers who want one consistent agent API across .NET, Python and Go, with first-class Azure/Microsoft Foundry integration
• Teams that want a genuinely open-source, self-hostable framework but also want the option of managed, scale-to-zero hosting without switching stacks

Key features

• Agent Harness: A production layer adding shell and filesystem access, human-in-the-loop approval flows, automatic context/token compaction and built-in FileMemory/FileAccess/Todo/Shell providers.
• Foundry Hosted Agents: Optional managed deployment with scale-to-zero infrastructure, per-session isolation and built-in OpenTelemetry/Application Insights observability.
• CodeAct: The model writes short Python programs to execute tool calls inside isolated Hyperlight micro-VMs instead of one-at-a-time tool calls — Microsoft’s own BUILD 2026 benchmarks report 52.4% faster completion and 63.9% fewer tokens on multi-step workloads.
• MCP + A2A protocol support: Native Model Context Protocol tool support and Agent2Agent protocol support for connecting to remote agents, aligning with the two dominant 2026 agent-interop standards.
• Graph-based multi-agent orchestration: Sequential, concurrent, handoff and group-collaboration workflow patterns for composing multiple agents.
• GitHub Copilot SDK integration: Agents can call into GitHub Copilot’s own shell and file-access tools directly as a provider.
• First-party AutoGen & Semantic Kernel migration guides: Official, documented migration paths for teams moving existing AutoGen or Semantic Kernel projects onto the unified framework.

What are Microsoft Agent Framework's use cases?

• Migrating an existing AutoGen or Semantic Kernel project: A team follows Microsoft’s first-party migration guide to move a production multi-agent system off a now-maintenance-mode framework onto the supported successor.
• Production agent that touches a real environment: An agent uses the Agent Harness’s shell/filesystem providers and human-in-the-loop approval to safely take real actions rather than only returning text.
• Cost- and latency-sensitive multi-step tool use: A workflow switches to CodeAct so the model batches tool calls into one executed program instead of many round trips, cutting tokens and latency on long tool-use chains.

What does Microsoft Agent Framework integrate with?

• Azure OpenAI, OpenAI, Microsoft Foundry, Anthropic, Ollama, Foundry Local, GitHub Copilot, Copilot Studio
• Model Context Protocol (MCP) tools
• Agent2Agent (A2A) protocol
• Google Gemini (Go SDK)
• Azure Functions & durable hosting
• OpenTelemetry / Application Insights

Why use Microsoft Agent Framework?

• Microsoft’s own supported path forward — AutoGen and Semantic Kernel are both explicitly in maintenance mode, with new capability development landing here first.
• Genuinely multi-provider despite the Microsoft branding: first-class Azure OpenAI and Foundry, plus OpenAI, Anthropic Claude, local models via Ollama, and Google Gemini through the Go SDK.
• A rare framework in this category that ships a real production-ops layer (Agent Harness, Hosted Agents, OpenTelemetry) rather than leaving deployment entirely to you.
• MIT-licensed and free to self-host at the core, with an optional managed hosting layer rather than a mandatory one.
• CodeAct’s measured token/latency reduction and native MCP + A2A support align it with where the rest of the 2026 agent ecosystem is standardizing.

Pros & cons

Microsoft Agent Framework pricing

• Open source (self-host)Free
• Foundry Hosted AgentsAzure/Foundry consumption

See current pricing on learn.microsoft.com ↗Compare Microsoft Agent Framework alternatives →

Microsoft Agent Framework review

Microsoft Agent Framework is what AutoGen and Semantic Kernel became once Microsoft decided to stop maintaining two separate agent stacks and support one instead. It reached a production-ready 1.0 GA on 2026-04-03, ships consistent APIs across .NET, Python and Go, and is genuinely more open than its Microsoft branding suggests — Anthropic’s Claude and local Ollama models are first-class citizens, not afterthoughts. What sets it apart from most frameworks in this category is that it doesn’t stop at orchestration: the Agent Harness gives an agent a real, guarded environment to act in, and Foundry Hosted Agents offers a managed path to production without forcing you off the free, self-hostable core. CodeAct’s measured token and latency savings and native MCP/A2A support show it building for where the rest of the ecosystem is heading, not just where Microsoft wants it to go. The honest catch is age: as a unified product it is roughly a quarter old, with a GitHub community still far smaller than LangChain’s or legacy AutoGen’s, and Microsoft’s own terms put the burden of risk on you for any non-Microsoft model or MCP server you plug in. Pick it if you are migrating off AutoGen or Semantic Kernel, or starting fresh with real Azure/Foundry investment; if neither applies, LangChain’s longer track record and wider ecosystem are still the safer default.

Microsoft’s unified, production-ready successor to AutoGen and Semantic Kernel (1.0 GA, April 2026) — genuinely multi-provider and ops-minded (Agent Harness, Foundry Hosted Agents, CodeAct), but young as a single product with a GitHub community still far smaller than its more established rivals.

Frequently asked questions