Skip to content

Harness profile

Microsoft Agent Framework: the successor to AutoGen and Semantic Kernel

Microsoft Agent Framework is Microsoft’s open-source framework for agents and multi-agent workflows in .NET and Python, and the direct successor to both AutoGen and Semantic Kernel. It is best for teams on Azure or .NET that want typed, checkpointed workflows with approvals and OpenTelemetry tracing.

StudiedLast reviewed 4 min read

Explained from public sources. Mai has not built this yet.

On this page

In plain English

Microsoft used to have two agent projects: AutoGen, a research framework for getting several agents to talk to each other, and Semantic Kernel, a toolkit for putting AI into business software. Microsoft Agent Framework merges them into one product, built by the same teams. You can build a single agent that uses tools, or a workflow where several agents and ordinary functions pass work along a fixed path, pause for a person, and pick up again later. It runs in your own .NET or Python code and works with Azure and with other model providers.

How it works

The documentation groups the framework into four areas: agents (a model with instructions, tools and MCP servers), a harness agent with planning, context compaction, file access and tool approvals for long tasks, workflows that connect agents and functions along explicit paths, and integrations with model providers, memory stores, middleware and evaluation services. Underneath sit model clients, an agent session for state, context providers for memory, middleware that can intercept what an agent does, and MCP clients for tools.

Workflows are graphs. Executors receive input, do work and emit output, and edges route messages between them, including conditional routes and fan-out to parallel branches that fan back in. Messages are typed, so the graph can be checked before it runs. Progress is checkpointed at the boundary of each step of the graph, so a long-running process can survive an interruption and resume. A workflow can also be wrapped so that it looks like a single agent to the rest of your code. Python additionally offers an experimental functional style, written with ordinary loops and branches instead of edges.

In: Request. Then Intake agent. Then Parallel: Pricing executor, Contract agent. Then Request for input. Then Tools and MCP servers. Then Checkpoints, OpenTelemetry traces.

A workflow routes typed messages between executors, checkpoints each step and pauses for a person

Multi-agent patterns. The 1.0 release ships ready-made orchestrations: sequential, concurrent, handoff, group chat and Magentic-One. They support streaming, checkpointing, human approvals and pause and resume.

Human-in-the-loop. Inside a workflow, a dedicated request step (or a request_info call in the functional style) stops the run, sends a question out to your application, and continues when the answer comes back.

Models and tools. The 1.0 announcement lists connectors for Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama. Agents call function tools and MCP servers. Memory can come from Foundry Agent Service, Mem0, Redis, Neo4j or your own store.

Observability. The repository describes built-in OpenTelemetry integration for distributed tracing, monitoring and debugging.

Languages and status. Version 1.0 was announced on 3 April 2026 for .NET and Python, with stable APIs. A Go version is in public preview in a separate repository. At 1.0, DevUI (a local browser debugger), Foundry hosted agents, skills and the agent harness were listed as preview features. The framework is MIT licensed. Agents and workflows can also be defined declaratively in YAML.

What it does best

  • One path forward from two projects. Teams on AutoGen or Semantic Kernel get an official successor and migration guides, rather than having to choose a new vendor.
  • .NET and Python as equals. Both reached 1.0 together, which suits organisations whose systems are written in C#.
  • Explicit, typed workflows. Typed messages, conditional edges and fan-out make the process inspectable and checkable before it runs.
  • Built for long-running work. Checkpoints and request steps let a workflow wait for a person or survive a restart.
  • Standard telemetry. OpenTelemetry means traces can go to whatever monitoring the organisation already uses.
  • Broad provider support. It is not limited to Azure models.

Where it stops

  • It leans towards Azure. The quick starts, hosting options and several preview features centre on Microsoft Foundry and Azure. Other providers work, but the deepest integration is with Microsoft’s own services.
  • Parts are still preview. Some features named in the docs were preview at 1.0, and their APIs may change.
  • Feature parity varies. The Go version lacks declarative agents, functional workflows and some other features, and the docs ask you to check provider-specific limitations.
  • Responsibility stays with you. Microsoft’s docs state plainly that using third-party models or servers is at your own risk, and that safety measures, data boundaries and testing are the builder’s job.

Where it goes wrong

The main risk is migration rather than the framework. Moving from AutoGen or Semantic Kernel is a migration with its own guides, not a version bump, and AutoGen itself is no longer maintained or recommended for production, so a team that delays moving is left on an unsupported base. The second is reaching for a multi-agent orchestration when one agent, or plain code, would do. Microsoft’s own guidance says that if you can write a function to handle the task, you should do that instead of using an agent.

How Mai AI uses it

I would use Microsoft Agent Framework when a client already runs on Microsoft: .NET services, Azure hosting, Microsoft 365 and Teams. In that setting its typed workflows, checkpoints and OpenTelemetry traces fit the tools their IT team already supports, and keeping data inside their Azure boundary is simpler. For a firm still on AutoGen or Semantic Kernel, it is the natural migration target.

The idea I take from it is Microsoft’s own rule: use an agent only where the task is open-ended, and a workflow with explicit steps everywhere else. That rule shapes how I design long-running jobs and where I place human approval gates. To see how these choices come together for one business, read how Mai AI designs custom AI operations systems.

Questions and answers

What happened to AutoGen?

Microsoft has replaced it. The AutoGen repository now says the project is in maintenance mode and no longer maintained, and points new and existing users to Microsoft Agent Framework, which it calls the successor.

Is Semantic Kernel being replaced by Microsoft Agent Framework?

Yes. Microsoft describes Agent Framework as the direct successor to both Semantic Kernel and AutoGen, built by the same teams, and publishes migration guides from each.

Is Microsoft Agent Framework production ready?

Version 1.0 was announced on 3 April 2026 for .NET and Python with stable APIs. Some features, such as DevUI and parts of the Foundry integration, were released as preview and may still change.

Is Microsoft Agent Framework free and does it only work with Azure?

The framework is open source under the MIT licence. It works with Azure OpenAI and Microsoft Foundry, and also with OpenAI, Anthropic, Amazon Bedrock, Google Gemini, Ollama and others. You pay for the models and services you call.

Sources

  1. Microsoft Agent Framework overviewlearn.microsoft.com
  2. Workflow conceptslearn.microsoft.com
  3. Microsoft Agent Framework version 1.0 announcementdevblogs.microsoft.com
  4. agent-framework on GitHubgithub.com
  5. autogen on GitHubgithub.com
  6. semantic-kernel on GitHubgithub.com