Blog | November 18, 2026 | MultiAgent Compare editors

How to migrate from AutoGen to Microsoft Agent Framework

Reviewed Nov 2026

Follow Microsoft's AutoGen migration guide, which covers model client setup, then single-agent features, then multi-agent orchestration. Turn each AssistantAgent into an Agent, rewrite FunctionTool wrappers as @tool functions, and, because Agent keeps no history between invocations, use AgentSession to manage conversation history when a conversation continues. Rebuild teams on the built-in orchestration patterns, which include Group Chat and Magentic. Microsoft says AutoGen has been archived and is no longer actively maintained.

What does Microsoft say about AutoGen's status?

Microsoft no longer maintains AutoGen and points its users to Agent Framework. The repository says Microsoft is not providing bug fixes, security fixes, new features, or support for the project. Elsewhere on the same page it describes AutoGen as in maintenance mode, with contributions limited to bug fixes, security patches and documentation improvements. The AutoGen docs site says existing users are encouraged to migrate using the migration guide.

Microsoft's overview calls Agent Framework the direct successor, created by the same teams. The AutoGen repository describes Agent Framework as a production-ready release with stable APIs and a commitment to long-term support, while for AutoGen itself Microsoft says it continues to advise against production use. For feature-level differences, see the AutoGen and Microsoft Agent Framework comparison.

What maps to what between AutoGen and Agent Framework?

Single-agent concepts carry over, while defaults, orchestration and runtime differ. Microsoft's guide lays out a practical migration path that starts with what stays the same and what changes.

  • Both libraries support function-style tools, token streaming, multimodal content, and async I/O.
  • AssistantAgent is single-turn unless you increase max_tool_iterations, whereas Agent is multi-turn by default and keeps invoking tools until it can return a final answer.
  • AutoGen wraps functions with FunctionTool, and Agent Framework uses @tool, infers schemas automatically, and adds hosted tools such as a code interpreter and web search.
  • Teams become workflows. The guide says AutoGen pairs an event-driven core with a high-level Team, while Agent Framework centers on a typed, graph-based Workflow that routes data along edges and activates executors when inputs are ready.
  • Distributed runtime code needs a plan. AutoGen offers embedded and experimental distributed runtimes, and Agent Framework focuses on single-process composition today, with distributed execution planned.
  • OpenAIAssistantAgent users should migrate to the Responses client for direct OpenAI or Azure OpenAI work, or use FoundryAgent when they need a service-managed agent.

How do model clients and tools change?

Start with the clients, since every agent takes one. In the guide's Agent Framework code, the OpenAI client reads the API key from the environment, and Azure OpenAI uses the same OpenAIChatCompletionClient class with explicit Azure routing inputs. For the Responses API, OpenAIChatClient provides Responses API support for direct OpenAI and Azure OpenAI routing, including reasoning models and structured responses not available in AutoGen.

Plain function tools port with the @tool change. Agents that you exposed to other agents as tools need their own pass. Microsoft's tools page says you can use an agent as a function tool for another agent, and the inner agent is converted to a function tool that the outer agent can call. The Python sample passes weather_agent.as_tool() to the main agent. For MCP servers, the same page says Local MCP Tools work with any provider that supports function tools.

How do teams and group chats become workflows?

Map each team to the closest built-in orchestration before you hand-build a graph. The guide has a Workflow versus GraphFlow overview and sections on RoundRobinGroupChat and MagenticOneGroupChat.

For a round-robin team, the group chat page imports GroupChatBuilder and shows how to build a group chat with custom speaker selection logic. Its orchestrator can use various strategies, such as round-robin, prompt-based or custom logic, to pick speakers. In a group chat, agents don't share the same session instance; the orchestrator synchronizes each agent's session with the conversation history before each turn. In Python, non-user messages retain text only in that shared history, and function calls, function results and approval payloads are filtered before broadcast.

If you ran Magentic-One, the Magentic orchestration is designed based on Magentic-One, which AutoGen invented. Microsoft also warns that it is untested how well the Magentic orchestration performs outside the original Magentic-One design.

Microsoft's workflow concepts page says all SDKs support graph-based workflows, and Python also offers an experimental functional workflow API. If you are deciding between agents passing control and agents sharing one conversation, read handoffs versus shared rooms.

What changes with sessions, checkpoints and human input?

Conversation state moves out of the agent. The guide says Agent is stateless and doesn't keep conversation history between invocations, unlike AssistantAgent, so you pass an AgentSession to continue a conversation. Its examples include Redis-backed sessions for persisting conversation state externally. Microsoft warns that reusing a session with a different agent configuration or provider can lead to invalid context.

Checkpoints save the state of a workflow at specific points so you can resume later, and they are created at the end of each superstep. One setup on the checkpoint page has a CheckpointManager provided when running the workflow, and another has a CheckpointStorage provided when creating a workflow. The built-in storage implementations are InMemoryCheckpointStorage for tests and demos, FileCheckpointStorage on local disk, and CosmosCheckpointStorage in the agent-framework-azure-cosmos package for production and cross-process workflows.

Built-in orchestration builders accept an optional name that groups their checkpoints, and Microsoft says to set an explicit name when several workflows of the same orchestration type share checkpoint storage.

Human input becomes a workflow request. When an executor calls ctx.request_info(), the workflow emits a request_info event, and the framework automatically routes the response to the executor's @response_handler method. Microsoft notes that sequential, concurrent, and group chat orchestrations do not pause for free-form user input on their own. Pending requests are saved as part of the checkpoint state and re-emitted when you restore.

In what order should I migrate?

Work from the bottom of the stack up, and keep each step runnable before you start the next one.

  • List the AutoGen pieces your project uses, such as AssistantAgent, teams and custom runtime code, and decide early what to do about anything that relies on the distributed runtime.
  • Install the Python SDK with pip install agent-framework. Microsoft notes that Agent Framework does not automatically load .env files, so call load_dotenv() at startup or set variables in your shell.
  • Port single agents and their tools, and review tool loops on purpose, because Agent is multi-turn by default.
  • Add an AgentSession wherever your AutoGen agent relied on remembered history.
  • Replace each team with the nearest built-in orchestration, then rebuild any remaining custom flow as a workflow graph.
  • Add checkpoint storage, starting with file storage locally. For sensitive tools, the handoff page shows @tool(approval_mode="always_require") for operations that need human approval.
  • Turn on tracing. Agent Framework emits traces, logs, and metrics according to the OpenTelemetry GenAI Semantic Conventions.

Is it safe to keep AutoGen running while I migrate?

Microsoft warns that without maintenance, vulnerabilities in this project and its dependencies may remain unaddressed. Weigh that risk for any service that handles sensitive data, and move those parts first.

Which languages can I target with Agent Framework?

The FAQ page says Agent Framework currently supports .NET (C#) and Python. The overview adds that Agent Framework for Go is in public preview, requires Go 1.26 or later, and does not implement declarative agents, CodeAct or functional workflows.

Can I migrate one agent at a time?

You can port and test one model client and one agent before you touch any team. Keep the remaining AutoGen teams in place until each of their agents has a working replacement.

Where can I compare AutoGen with other frameworks before I commit?

Use the AutoGen and CrewAI comparison if you are weighing a target other than Microsoft's own successor.

Where do these facts come from?

Reviewed Nov 2026