Migrating from AutoGen or Semantic Kernel to Microsoft Agent Framework

What is it?

If your AutoGen agent is "a model, some instructions and a few tools", moving it is mostly renaming. If you built group chats or custom event loops, you will redesign them as workflows: nodes that pass typed data along edges. Semantic Kernel users have an official guide too. Either way the safe approach is the same: port one agent at a time, keep a test that checks its behaviour, and cut over only when old and new agree.

Also known as: AutoGen to Agent Framework migration, Semantic Kernel to Agent Framework migration, migrate AutoGen, migrate Semantic Kernel, AutoGen migration guide, AssistantAgent to Agent, GraphFlow to Workflow, AutoGen vs Microsoft Agent Framework.

Technical definition

Microsoft's AutoGen guide (Python SDK) states what stays the same — agents wrap a model client, take instructions and tools, and both libraries support function tools, token streaming, multimodal content and async I/O — and lists key differences: (1) orchestration — AutoGen pairs an event-driven core with a high-level Team, whereas Agent Framework centres on a typed, graph-based Workflow that routes data along edges and activates executors when inputs are ready; (2) tools — AutoGen's FunctionTool becomes the @tool decorator with automatically inferred schemas, plus hosted tools such as a code interpreter and web search; (3) agent behaviour — AutoGen's AssistantAgent is single-turn unless max_tool_iterations is raised, while Agent is multi-turn by default and keeps calling tools until it can answer; (4) runtime — AutoGen offered embedded and experimental distributed runtimes, while Agent Framework focuses on single-process composition (distributed execution is described as planned). Conversation state moves to an AgentSession object. GraphFlow versus Workflow: control-flow edges become data-flow edges, nodes can be agents, functions or sub-workflows, and typing is strong. The guide also covers MCP tools (stdio, streamable HTTP, websocket), agent-as-a-tool, middleware, human-in-the-loop via a request/response API, checkpointing and resuming, and OpenTelemetry. The Semantic Kernel README points to its own migration guide; this page does not restate that guide because its contents could not be extracted at check time.

Entity type: technique · Freshness class: volatile.

When should I use it?

Plan a migration if you rely on AutoGen or Semantic Kernel for a system that must stay supported; both repositories now recommend MAF for new work. A stable legacy system may reasonably stay put until you have capacity and tests.

Implementation patterns

Strangler migration: run old and new agents side by side on the same test inputs and compare outputs.
Port tools first (they are the most mechanical), then single agents, then orchestration.
Replace broadcast group chats with explicit edges so data flow is visible and testable.
Keep an evaluation set of real tasks and re-run it after each ported component (Agent Evaluation).

Technologies and frameworks

Commonly associated: AssistantAgent mapping, FunctionTool mapping, GraphFlow mapping, AgentSession, MCP tools, OpenTelemetry. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

Inventory agents, tools, teams, runtimes and model clients in the old codebase.
Write or collect representative tasks and expected behaviours.
Port model clients and tools; confirm tool schemas match.
Port each single agent; compare outputs with the old agent on the test set.
Redesign multi-agent orchestration as workflows; add checkpoints and human approval where the old system lacked them.
Add tracing, then cut over gradually and keep the old code until results match.

Common failure modes

Assuming AutoGen's single-turn default carries over: the new Agent loops on tools until it can answer, which changes cost and latency.
Porting a group chat line-for-line instead of redesigning as a workflow.
Relying on the experimental distributed runtime without a plan: the guide describes single-process composition.
Trusting the guide's "future patterns" note as current: it lists Swarm and SelectorGroupChat as in development while the framework README lists handoff and group collaboration as supported — check the repository.
Migrating on a deadline with no behaviour tests.

Troubleshooting

Agent now makes many more tool calls — Agent is multi-turn by default; set limits and review tool descriptions.
Lost conversation memory — Pass an AgentSession across calls; without a session each call is independent.
Pattern you used in AutoGen has no direct equivalent — Check the guide's pattern sections and the repository samples; otherwise rebuild it as a workflow graph.

Contrasted with

AutoGen (Microsoft) — The source framework (maintenance mode).
Semantic Kernel — The other source framework; Microsoft publishes a separate guide.
Microsoft Agent Framework — The target.

Questions this page answers

How do I migrate from AutoGen to Microsoft Agent Framework?
How do I migrate from Semantic Kernel to Microsoft Agent Framework?
What replaces AutoGen's AssistantAgent and Team?
What is the difference between AutoGen GraphFlow and Agent Framework Workflow?
Should I migrate an existing AutoGen project?

Typical goals: port existing AutoGen code; port existing Semantic Kernel code; estimate the effort and risks of migrating.

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):

What stays the same and the four key differences (orchestration style, tools, agent turn behaviour, runtime) as stated in the AutoGen migration guide.
GraphFlow versus Workflow comparison table (control flow vs data flow; agents-only vs agents/functions/sub-workflows; typing).
The guide covers MCP tools, agent-as-a-tool, middleware, human-in-the-loop, checkpointing and observability.
The guide's "Future Patterns" note lists Swarm and SelectorGroupChat as in development, while the MAF README lists handoff and group collaboration among supported patterns (an inconsistency, reported rather than resolved).
The Semantic Kernel README links to a Semantic Kernel migration guide and to an Agent Framework blog post.

Primary sources consulted:

Not independently verified:

The contents of the Semantic Kernel migration guide (the page could not be read at check time), so no Semantic Kernel API mappings are given.
Whether the guide's code samples run unchanged on the current release.
The suggested migration steps and patterns above are general engineering practice, not Microsoft guidance.

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:

Which AutoGen patterns have first-class equivalents (the guide's roadmap note may be stale).
Whether distributed execution has shipped.
API and package names, which changed during the framework's development.

Further reading and canonical sources (not all checked):

AutoGen to Microsoft Agent Framework Migration Guide — learn.microsoft.com/en-us/agent-framework/migration-guide/from-autogen/
Semantic Kernel to Agent Framework migration guide — learn.microsoft.com/en-us/agent-framework/migration-guide/from-semantic-kernel
Semantic Kernel and Microsoft Agent Framework (Agent Framework blog) — devblogs.microsoft.com/agent-framework/semantic-kernel-and-microsoft-agent-framework/