EXPLAINER

Microsoft Agent Framework

frameworkfreshness: volatilekey claims checked 2026-10-09

Microsoft Agent Framework (MAF) is Microsoft's open-source framework for building agents and multi-agent workflows in .NET and Python, positioned by Microsoft as the successor to both AutoGen and Semantic Kernel.

Short answer: Microsoft Agent Framework (MAF) is an open, multi-language framework (Python and .NET, with a separate Go SDK) for production-grade agents and graph-based multi-agent workflows. The AutoGen and Semantic Kernel repositories both name it as their successor. Check the repository for the current version before you adopt it.

What is it?

Microsoft built two well-known agent toolkits: AutoGen (research-born, conversation-style multi-agent) and Semantic Kernel (enterprise SDK, strong in .NET). Microsoft Agent Framework is the single framework that replaces both. You build an agent from a model client, instructions and tools; for anything bigger than one agent you describe a workflow — a graph of steps that can include agents, plain functions and other workflows — with checkpoints, streaming and human approval steps built in.

Also known as: Microsoft Agent Framework, MAF Microsoft, Microsoft.Agents.AI, AutoGen successor, Semantic Kernel successor.

Technical definition

Microsoft's repository describes MAF as an open, multi-language framework for production-grade AI agents and multi-agent workflows in .NET and Python (the Go SDK lives in a separate repository). Documented capabilities include: multiple model providers (Microsoft Foundry, Azure OpenAI, OpenAI and the GitHub Copilot SDK among them); a middleware system for request/response processing; graph-based workflows with sequential, concurrent, handoff and group-collaboration patterns, plus checkpointing, streaming, human-in-the-loop and "time-travel"; built-in OpenTelemetry observability; declarative (YAML) agents; "agent skills"; a developer UI called DevUI; Foundry-hosted agents; and an experimental "AF Labs" package set. Install names given in the README are pip install agent-framework for Python and dotnet add package Microsoft.Agents.AI for .NET. The AutoGen and Semantic Kernel READMEs each describe MAF as their enterprise-ready successor, and the Semantic Kernel README says MAF reached version 1.0 as a production-ready release with stable APIs. Interoperability with What is MCP (Model Context Protocol)? and A2A (Agent2Agent) Protocol is stated in that README. Version numbers, provider lists and package names change; confirm them in the repository and Microsoft Learn.

Entity type: framework · Freshness class: volatile.

When should I use it?

Consider for new agent projects on the Microsoft stack, or for teams moving off AutoGen or Semantic Kernel. If you only need one prompt and a tool or two, a direct API call is simpler (AI Framework vs Direct Model API).

Implementation patterns

  • Single agent: a client, instructions and a few tools, run with the agent's run method.
  • Workflow graph: executors (agents, functions or sub-workflows) joined by edges that route typed messages.
  • Human-in-the-loop: pause a workflow on a request, resume it with the person's response.
  • Checkpoint long workflows so a failure resumes instead of restarting.
  • Add middleware for logging, content filtering and error handling around every model or tool call.
  • Expose tools through MCP servers rather than bespoke glue.

Technologies and frameworks

Commonly associated: Python SDK, .NET SDK, Microsoft Foundry, Azure OpenAI, OpenAI, OpenTelemetry, MCP, A2A, DevUI. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Write down whether you need one agent, a fixed pipeline, or dynamic multi-agent collaboration.
  2. Install the Python or .NET package and run the repository's quick-start sample against a test model deployment.
  3. Add tools one at a time; keep write actions behind approval (Permissions and Security for AI Integrations).
  4. Move to a workflow only when you need branching, parallelism, checkpoints or human approval.
  5. Turn on OpenTelemetry tracing early (LLM Observability (Tracing, Logging and Monitoring)).
  6. Pin versions and re-read the release notes before upgrading.

Common failure modes

  • Copying AutoGen's event-driven or Semantic Kernel's plugin code across unchanged — the programming model differs (see Migrating from AutoGen or Semantic Kernel to Microsoft Agent Framework).
  • Building a workflow for what a single agent with tools can do.
  • Granting an agent broad credentials; MAF does not remove the need for Permissions and Security for AI Integrations.
  • Treating tutorial snippets as stable: names and packages have changed between releases.
  • Assuming distributed execution exists because AutoGen had an experimental distributed runtime — the migration guide describes single-process composition.

Troubleshooting

  • Sample from a blog post does not run — Compare against the repository samples for your installed version; APIs and package names move.
  • Authentication to the model provider fails — Follow the provider-specific sample (for example Azure CLI sign-in for Foundry) and check the endpoint and model environment variables.
  • Workflow stalls — Check that every executor's output type matches the next executor's input and that human-input requests are answered.

Contrasted with

  • AutoGen (Microsoft) — AutoGen is in maintenance mode (no new features, community managed) and recommends MAF for new projects.
  • Semantic Kernel — Semantic Kernel's README announces MAF as its successor; existing Semantic Kernel code needs a migration plan.
  • LangGraph — LangGraph is an independent, graph-based agent runtime from LangChain; MAF is Microsoft's equivalent-scope framework with first-class .NET.
  • OpenAI Agents SDK — OpenAI's SDK is lighter and OpenAI-centred; MAF targets multiple providers and enterprise hosting.

Questions this page answers

  • What is Microsoft Agent Framework?
  • Is Microsoft Agent Framework the replacement for AutoGen and Semantic Kernel?
  • Does Microsoft Agent Framework support Python and .NET?
  • How do I build a workflow in Microsoft Agent Framework?
  • Is Microsoft Agent Framework production ready?

Typical goals: decide whether to build on Microsoft Agent Framework; understand how it relates to AutoGen and Semantic Kernel; build a first agent or workflow with it.

Sources and verification

Status: key claims checked on 2026-10-09 against the primary sources listed below; everything else is not independently verified. The claims under "Checked against primary sources" were compared with the linked pages on that date. Pages change, so re-check anything time-sensitive before relying on it.

Checked against primary sources (2026-10-09):

  • MAF is described as an open, multi-language framework for production-grade AI agents and multi-agent workflows in .NET and Python, with the Go SDK in a separate repository.
  • README feature list: multiple providers, middleware, graph workflows (sequential, concurrent, handoff, group collaboration) with checkpointing, streaming, human-in-the-loop and time-travel, OpenTelemetry, declarative YAML agents, agent skills, DevUI, AF Labs.
  • Install commands: pip install agent-framework; dotnet add package Microsoft.Agents.AI.
  • The AutoGen README states AutoGen is in maintenance mode and names MAF as the enterprise-ready successor.
  • The Semantic Kernel README states MAF is its enterprise-ready successor and that MAF reached version 1.0 as a production-ready release.
  • The migration guide states MAF is developed by the core AutoGen and Semantic Kernel teams.

Primary sources consulted:

Not independently verified:

  • The exact current release number and date of each SDK.
  • The licence.
  • Claims about adoption, performance or "enterprise" suitability, which are vendor positioning rather than testable facts.
  • Feature parity between the Python and .NET SDKs.

Freshness class: volatile — Products, versions, protocols, benchmarks, leaderboards, regulations and dates. Verify against the primary source before relying on specifics; re-review every 3 months.

Time-sensitive claims to verify before relying on them:

  • Current version number and release status of each language SDK (README was read on the check date).
  • Supported providers and hosting options.
  • Go SDK status (separate repository).
  • Whether distributed multi-process execution has shipped (the AutoGen migration guide says it is planned).
  • Licence terms (not checked).

Further reading and canonical sources (not all checked):

  • Microsoft Agent Framework repository README — github.com/microsoft/agent-framework
  • Microsoft Agent Framework documentation — learn.microsoft.com/agent-framework
  • AutoGen repository README (maintenance-mode notice) — github.com/microsoft/autogen
  • Semantic Kernel repository README (successor notice) — github.com/microsoft/semantic-kernel