Microsoft killed AutoGen. Not quietly, not gradually. In October 2025, the company announced that both AutoGen and Semantic Kernel would enter maintenance mode, their teams merged into a single unit building the new Microsoft Agent Framework. GA is targeted for the end of Q1 2026, with stable APIs, production SLAs, and compliance certifications (SOC 2, HIPAA) baked in.
If you have production code running on AutoGen 0.4, this is your migration signal. If you are evaluating agent frameworks for a new project, this changes the calculus. Here is what Microsoft actually built, what it replaces, and whether it deserves your attention next to LangGraph, CrewAI, and the rest.
Why Microsoft Merged Two Frameworks Into One
The honest answer: AutoGen and Semantic Kernel were solving overlapping problems for different audiences, and neither was complete on its own.
AutoGen (originally a Microsoft Research project) excelled at multi-agent patterns. Its conversational agent model let you spin up teams of specialized agents that could coordinate, debate, and iterate on problems. Researchers loved it. The AutoGen paper was cited thousands of times. But when enterprise teams tried to deploy it, they hit walls: no built-in session management, limited telemetry, no production deployment story, and a Python-only codebase that excluded .NET shops.
Semantic Kernel, meanwhile, was the enterprise SDK. Type-safe, multi-language (C#, Python, Java), deep Azure integration, filters, middleware, and telemetry out of the box. But its agent orchestration story was thin. Building multi-agent workflows required custom plumbing that AutoGen gave you for free.
Microsoft’s solution: take AutoGen’s multi-agent abstractions, layer them on Semantic Kernel’s enterprise foundation, and add a new workflow engine on top. Both original teams now work under one umbrella. Both original repos enter maintenance mode, receiving only bug fixes and security patches, with no new features.
According to VentureBeat’s reporting, the organizational merger is not cosmetic. It allocates substantial engineering resources to a single codebase, which means faster iteration and fewer “works in AutoGen but not Semantic Kernel” issues.
What Microsoft Agent Framework Actually Contains
The framework has two primary building blocks: Agents and Workflows.
Agents: The Individual Units
An Agent in Microsoft Agent Framework is an LLM-backed process that takes input, optionally calls tools or MCP servers, and produces output. So far, nothing different from any other framework.
What sets it apart is the stack of optional layers:
- Session management: Persists conversation state across multiple turns. This was one of the biggest gaps in standalone AutoGen.
- Context providers: Feed the agent memory, knowledge bases, or RAG results as context. Think of these as Semantic Kernel’s memory plugins, now integrated at the framework level.
- Middleware: Intercept any agent action before it executes. Useful for logging, guardrails, content filtering, or injecting compliance checks.
- MCP clients: Native Model Context Protocol support for tool integration, following the same MCP standard that Anthropic, OpenAI, and Google have adopted.
The supported model providers are Azure OpenAI, OpenAI, and Azure AI. No Anthropic or Google support out of the box, which is a significant limitation if you want multi-provider redundancy.
Workflows: The Orchestration Layer
Workflows are where the framework introduces genuinely new capability. Instead of relying purely on LLM-driven routing (the “let the model figure out what to do next” approach), workflows give you a graph-based execution engine with deterministic control.
from agent_framework import Workflow, Agent
workflow = Workflow()
researcher = Agent(name="researcher", model="gpt-4o")
writer = Agent(name="writer", model="gpt-4o")
# Define explicit execution paths
workflow.add_edge(researcher, writer)
workflow.add_conditional_edge(
writer,
condition=lambda output: "needs_revision" in output,
targets={True: researcher, False: "end"}
)
Key workflow features:
- Conditional routing: Branch execution based on agent output, not just LLM decisions.
- Parallel processing: Run independent agent tasks concurrently.
- Checkpointing: Save workflow state and resume later, critical for long-running processes.
- Human-in-the-loop patterns: Pause execution, wait for human input, and resume. This maps directly to what the EU AI Act Article 14 requires for high-risk systems.
- Multi-agent orchestration patterns: Sequential pipelines, concurrent fan-out, supervisor hierarchies, dynamic handoffs, and what Microsoft calls “Magentic” (a flexible pattern combining multiple approaches).
If you have read our multi-agent orchestration guide, these patterns will look familiar. The difference is that Microsoft provides them as first-class framework primitives rather than patterns you assemble yourself.
How It Compares to LangGraph and CrewAI
The comparison everyone wants. Here is the honest breakdown.
Microsoft Agent Framework vs. LangGraph
LangGraph remains the default recommendation for most teams, and the LangChain survey data backs this up: it has the largest production footprint, the most mature tooling (LangSmith for observability, LangGraph Platform for deployment), and the lowest latency in recent benchmarks.
Where Microsoft Agent Framework wins:
- Multi-language support: C#, Python, and Java. LangGraph is Python-first with JavaScript/TypeScript support via LangGraph.js. If your team is a .NET shop, this is not a minor detail; it is a dealbreaker in LangGraph’s direction.
- Azure integration: Native Azure AI Foundry deployment with built-in identity management (Entra ID), autoscaling, and governance. LangGraph Platform handles deployment too, but not with the same depth of cloud-native integration for Azure customers.
- Enterprise support contracts: Production SLAs, 24/7 support, compliance certifications. LangChain offers enterprise support through LangSmith, but Microsoft’s support story is more familiar to IT procurement teams.
- Governance built-in: Every agent gets an Entra identity with lifecycle management. Publishing an agent creates a managed Azure resource with audit trails, RBAC, and data isolation. LangGraph expects you to wire this up yourself.
Where LangGraph wins:
- Maturity: LangGraph has been production-tested at scale by Klarna, Replit, and hundreds of other companies. Microsoft Agent Framework is still in preview.
- Model flexibility: LangGraph works with any LLM provider. Microsoft’s framework officially supports only Azure OpenAI, OpenAI, and Azure AI models.
- Community and ecosystem: More tutorials, examples, integrations, and battle-tested patterns.
- Performance: Lower latency, lower token usage in head-to-head benchmarks.
Microsoft Agent Framework vs. CrewAI
CrewAI’s strength is speed to prototype. You can have a working multi-agent system in under 50 lines of Python. Microsoft Agent Framework requires more boilerplate but gives you more control over execution flow.
CrewAI remains the better choice for rapid prototyping, internal tools, and projects that do not need enterprise governance. Microsoft Agent Framework is the better choice if you need compliance, multi-language support, or are already invested in Azure.
The Migration Path from AutoGen
If you are running AutoGen 0.4.x in production, you need a migration plan. Microsoft has published migration guides and a DevUI for visualizing multi-agent workflows.
The core concepts map directly:
| AutoGen Concept | Agent Framework Equivalent |
|---|---|
AssistantAgent | Agent with model client |
UserProxyAgent | Workflow with human-in-the-loop edge |
GroupChat | Workflow with multi-agent orchestration pattern |
| Event-driven messaging | Workflow edges with conditional routing |
ConversableAgent | Agent with session management |
The biggest breaking change is the shift from AutoGen’s event-driven messaging model to the workflow graph model. In AutoGen, agents communicated through a message bus and the orchestration emerged from the conversation. In Agent Framework, you define explicit execution paths. This is more predictable but requires rethinking your agent interactions.
Microsoft states that AutoGen will continue receiving bug fixes and security patches, so there is no immediate deadline. But new features will only land in Agent Framework.
Who Should Actually Use This
Use Microsoft Agent Framework if:
- Your organization runs on Azure and .NET. The integration depth is unmatched.
- You need compliance certifications (SOC 2, HIPAA) from your framework vendor, not just from your cloud provider.
- You want agent identity management handled by Entra ID rather than building it yourself.
- You are building long-running, human-in-the-loop workflows that need checkpointing and state persistence.
Stick with LangGraph if:
- You want the largest community, most examples, and most battle-tested deployment patterns.
- You need multi-provider model support (Anthropic, Google, Mistral, open-source models).
- Your team works primarily in Python and does not need .NET or Java support.
- You are already using LangSmith for observability and evaluation.
Choose CrewAI if:
- Speed to prototype matters more than enterprise governance.
- Your agents are relatively straightforward and do not need complex graph-based workflows.
The framework landscape in 2026 is consolidating. Google has Vertex AI Agent Builder. OpenAI has the Agents SDK. Amazon has Bedrock Agents. Microsoft now has Agent Framework. The era of picking a framework from a startup and hoping they survive long enough to maintain it is shifting toward platform-backed options with enterprise support. That does not make the indie frameworks worse; it just means the decision criteria now include vendor viability alongside technical capability.
Frequently Asked Questions
What happened to Microsoft AutoGen?
Microsoft retired AutoGen as an active development project in October 2025, merging it with Semantic Kernel into the new Microsoft Agent Framework. AutoGen continues to receive bug fixes and security patches but no new features. The teams behind both AutoGen and Semantic Kernel now work together on Agent Framework.
When will Microsoft Agent Framework reach GA?
Microsoft targets Agent Framework 1.0 general availability by the end of Q1 2026, with stable versioned APIs, production SLAs, and compliance certifications including SOC 2 and HIPAA.
Should I migrate from AutoGen to Microsoft Agent Framework?
Yes, if you plan to continue building on a Microsoft-backed framework. AutoGen is in maintenance mode with no new features planned. Microsoft has published migration guides and the core concepts map directly. The biggest change is moving from event-driven messaging to graph-based workflows.
How does Microsoft Agent Framework compare to LangGraph?
LangGraph has more production maturity, broader model support, and a larger community. Microsoft Agent Framework offers better Azure integration, multi-language support (C#, Python, Java), built-in enterprise governance, and compliance certifications. LangGraph remains the better default choice unless you are deeply invested in Azure and .NET.
What programming languages does Microsoft Agent Framework support?
Microsoft Agent Framework supports C#/.NET, Python, and Java. This multi-language support is one of its key differentiators, particularly for enterprise teams that work primarily in .NET rather than Python.