Between them, AutoGen and Semantic Kernel accumulated more than 75,000 GitHub stars across thousands of enterprise deployments. On April 3, 2026, Microsoft told both communities to start planning a move. The new destination is called Microsoft Agent Framework, and the "successor" label is accurate - but it does not mean what a lot of teams are assuming it means.
What MAF 1.0 actually bundles
For three years, Microsoft ran parallel bets. Semantic Kernel handled model connectors, memory, and tool calling. AutoGen handled multi-agent conversation patterns. Teams had to pick one, or stitch them together with custom glue. That friction was loud and well-documented in community channels. Developers complained that starting a new project meant reading the same release notes twice and picking whichever framework had the feature they needed that week.
MAF is the direct successor to both, merging AutoGen's agent abstractions with Semantic Kernel's enterprise features. Microsoft shipped version 1.0 on April 3, 2026, with stable APIs, native MCP and A2A support, and a long-term support commitment.
The actual capability set at 1.0:
Unified agent abstraction. A single
AIAgentprimitive that works across model providers, plus a "Harness Agent" designed for long, multi-step autonomous tasks.Graph-based orchestration. The multi-agent programming model shifted from AutoGen's conversation-centric design - where agents talk to each other in chat threads - to a graph-based model where agents are nodes in a directed graph with explicit transitions.
Protocol support. Full integration with the Model Context Protocol lets agents dynamically discover and invoke external tools and data sources without manual integration code. On A2A, the picture is blurrier: the 1.0 post reiterated MCP support and said A2A 1.0 support is coming soon, positioning those protocols as part of its cross-runtime agent story
which means you should verify the actual A2A status against the MAF devblog before you architect around it, not just trust a third-party summary.
Six model providers, one-line swap. Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, Ollama - swapping providers is a one-line change.
Observability baked in. State and memory: sessions, threads, context providers, and checkpointing ship as first-class primitives rather than community recipes.
Over 12,000 organizations were already using Semantic Kernel in production before the merge, and the framework handles more than 1 billion AI interactions monthly across Azure services. That installed base is the reason Microsoft committed to LTS on the 1.0 APIs - they cannot afford to break it.
What the migration actually costs
This is where the marketing and the reality split.
"Successor" does not mean "drop-in replacement." Semantic Kernel encouraged a kernel-plus-plugin mental model. AutoGen encouraged conversational agents and group chat patterns. Agent Framework normalizes agents, workflows, tools, state, durability, and hosting behind one model.
From Semantic Kernel: The migration is manageable.
The kernel, plugin model, and connector system are structurally the same. You update import paths and replace deprecated class names.
The recommended order is
plugins first, then the kernel layer, then agents - plugin code is largely compatible, and the bigger work is moving from the old Kernel object to the new agent primitives.
One specific gotcha:
if any non-agent code path returns Semantic Kernel's ChatMessageContent - a logging helper, a DTO mapper, a RAG pipeline - it will not compile against the agent layer anymore. Agent Framework uses Microsoft.Extensions.AI ChatMessage. Grep for ChatMessageContent across the whole solution before you call the migration done.
From AutoGen: Harder.
The group chat and handoff patterns map cleanly, but the new workflow engine is a more opinionated model than AutoGen's freeform GroupChatManager.
Breaking changes are real. Session state has a new API. Filter pipelines are unified. Telemetry tags were renamed for OTel alignment. Teams with large codebases should plan a dedicated migration sprint rather than treating this as a drop-in upgrade.
| Migration path | Effort | What breaks |
|---|---|---|
| Semantic Kernel → MAF | 2-4 hours for a typical app | ChatMessageContent type, AgentGroupChat preview class, FunctionChoiceBehavior.Auto() |
| AutoGen → MAF | Dedicated sprint | Entire orchestration model (conversation → graph), AssistantAgent, GroupChatManager, import paths |
| New project | Start here | - |
Microsoft committed in October 2025 to keep supporting Semantic Kernel v1.x, fixing critical bugs and security issues, for at least one year after Agent Framework became generally available
- which puts the earliest end of that window in April 2027. AutoGen moved to maintenance mode on October 2, 2025: it receives critical bug fixes and security patches, no new features, and is community managed going forward.
Beagle in action#engineering, 10:02amWhat is genuinely new versus incremental
The consolidation is real, but it is worth being precise about what MAF adds versus what it merely reorganises.
Genuinely new:
- LangGraph 1.0 reached GA in October 2025, making checkpointing, durable execution, and human-in-the-loop approvals first-class primitives rather than community recipes.
MAF's 1.0 does the same thing for the Microsoft stack - durable, checkpointed execution was previously experimental in both predecessor frameworks.
The new orchestration patterns - sequential pipelines, handoff delegation, human-in-the-loop approval, MCP integration - exist only in the new framework. These were not available in the stable SK or AutoGen releases.
In April 2026, Microsoft positioned the framework as production-ready with stable APIs, long-term-support commitment, MCP support, migration assistants, DevUI, Foundry integration, observability, evaluations, frontend streaming adapters, skills, and agent harness support.
Incremental:
- The model provider abstraction existed in Semantic Kernel already. MAF extends it to six providers cleanly, but the concept is not new.
- Graph-based orchestration has been the norm in LangGraph for over a year. MAF is catching up to a pattern the open-source community established, not pioneering it.
It exists but will be deprecated.
The part vendors do not emphasize
Per the Microsoft Agent Framework devblog dated 3 April 2026, A2A 1.0 support is listed as coming soon rather than shipped. Native MCP support means the same agent can call MCP tools without a custom bridge. Treat the A2A "coming soon" note as a to-verify item, not a promise you can build against yet.
That matters for teams planning cross-framework agent coordination. If you are building a workflow where a MAF agent needs to coordinate with a LangGraph agent via A2A today, verify the current state of the A2A implementation against the official devblog before committing to that architecture.
The second thing to check is the AgentSession cleanup behavior.
In Semantic Kernel, thread.DeleteAsync() existed as a self-contained method. In Agent Framework, there is no session.DeleteAsync() - not all providers support hosted session deletion. If you use a provider that supports it (OpenAI Assistants), you are responsible for tracking session IDs and calling the deletion through the provider SDK. This is a common memory leak in early migrations.
Teams using a teammate like Beagle to triage engineering questions in Slack already have a pattern for this kind of docs-archaeology work - the answer to "which AutoGen patterns need rewrites?" exists in primary sources, it just takes time to find and synthesize.
Microsoft Agent Framework: common questions
What is Microsoft Agent Framework and what does it replace?
Microsoft Agent Framework reached 1.0 for Python and .NET on April 2, 2026. Microsoft describes it as the direct successor to both AutoGen and Semantic Kernel, combining AutoGen's agent abstractions with Semantic Kernel's state management, middleware, and telemetry. AutoGen is in maintenance mode; Semantic Kernel gets critical patches through at least April 2027.
How hard is migrating from AutoGen to Microsoft Agent Framework?
Harder than the marketing implies. The multi-agent programming model shifted from AutoGen's conversation-centric design to a graph-based model. If you have AutoGen code in production, migration requires real work. A compatibility layer exists, but it will be deprecated. Plan a dedicated sprint for any non-trivial AutoGen codebase.
Does Microsoft Agent Framework support MCP and A2A natively?
MCP support is native at 1.0. MCP lets agents dynamically discover and invoke external tools exposed over MCP-compliant servers. A2A 1.0 support - enabling cross-runtime agent collaboration - is listed as coming soon. Verify the current A2A status against the official MAF devblog before designing a cross-framework agent architecture around it.
Should new projects use Microsoft Agent Framework or LangGraph?
Depends on your stack. LangGraph supports sequential, concurrent, handoff, and group patterns with checkpointing, streaming, human-in-the-loop, and time travel, with OpenTelemetry built in. If you want stateful graphs in Python, LangGraph reached 1.0 in October 2025. If you are on .NET or migrating from AutoGen or Semantic Kernel, Agent Framework gives you Python and .NET at 1.0 since April 2026.
Is Semantic Kernel deprecated?
Semantic Kernel is not switched off. Microsoft committed to keep supporting Semantic Kernel v1.x, fixing critical bugs and security issues, for at least one year after Agent Framework became generally available, which puts the earliest end of that window in April 2027. The majority of new features will be built for Agent Framework. Existing Semantic Kernel projects can stay put for now, but new features will not arrive there.