AutoGen, AG2, Semantic Kernel and Microsoft Agent Framework: What Happened and What to Use

The short version

Microsoft had two well known open source projects for building agents and model powered applications: AutoGen, known for multi-agent conversations, and Semantic Kernel. In October 2025 Microsoft introduced Microsoft Agent Framework, which brings ideas from both into one framework, and in April 2026 Microsoft published its 1.0 release for .NET and Python, describing it as production ready with stable APIs and a commitment to long term support.

If you are starting a new project on Microsoft's stack today, the official guidance points to Microsoft Agent Framework. If you maintain an AutoGen or Semantic Kernel codebase, Microsoft publishes migration guides for both. The sections below give the status of each project as stated in its own repository or announcement, because tutorials written before 2026 often describe a situation that no longer exists.

Status of each project

Check the date on any tutorial you read, and check the repository README before you add a dependency. The list below reflects what the official sources said when this course was written. Semantic Kernel's support statement is from Microsoft's October 2025 blog post, which said support for Semantic Kernel v1.x continues for at least one year after Microsoft Agent Framework reaches general availability, with critical bugs and security issues still addressed, while the majority of new features go into Agent Framework.

• AutoGen (microsoft/autogen): the README states the project is archived, no longer receives bug fixes, security fixes or features from Microsoft, and that new users should start with Microsoft Agent Framework. Code is MIT licensed.
• Semantic Kernel: still supported for critical bugs and security issues, with most new features going to Agent Framework.
• Microsoft Agent Framework (microsoft/agent-framework): MIT licensed, Python and .NET, with a Go version in a separate repository.
• AG2 (ag2ai/ag2): a separate, community run project that describes itself as formerly AutoGen. Python 3.10 or later, Apache 2.0 licensed, with original AutoGen code under MIT.

Core concepts in Microsoft Agent Framework

Agent Framework has two layers. Agents are single model driven units with instructions, tools, middleware and memory or context providers. Workflows are graph based and support multi-agent orchestration patterns that Visual Studio Magazine's 1.0 coverage lists as sequential, concurrent, handoff, group chat and Magentic-One. The repository README adds checkpointing, streaming, human in the loop control, declarative agents in YAML and built in OpenTelemetry observability.

It is not tied to one model vendor. The 1.0 release included connectors for Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama, and MCP support. Some pieces, such as the browser based DevUI and certain hosting integrations, were still marked preview at 1.0, so check the status of each feature you depend on before you design around it, and pin package versions in your requirements file.

A first agent in Python

The Microsoft Learn quick start installs agent-framework-foundry and azure-identity with pip, signs in with az login, and creates an agent backed by a Microsoft Foundry project. The code below follows that quick start with placeholder values. Replace the endpoint and model with your own project's deployment name. The docs warn that convenient development credentials should be replaced with a specific credential, such as a managed identity, in production.

If you do not use Azure, the agent-framework metapackage installs the core together with many optional integrations; check the docs for the client class that matches your provider, because each connector has its own setup. Notice that the agent API is async, so in a web service such as FastAPI you would await agent.run inside your request handler rather than calling asyncio.run.

# Illustrative, check the official docs for the current API
# Based on learn.microsoft.com/agent-framework/tutorials/quick-start
# pip install agent-framework-foundry azure-identity
import asyncio

from agent_framework import Agent
from agent_framework.foundry import FoundryChatClient
from azure.identity import AzureCliCredential


async def main() -> None:
    agent = Agent(
        client=FoundryChatClient(
            project_endpoint="https://your-account.services.ai.azure.com/api/projects/your-project",
            model="your-model-deployment-name",
            credential=AzureCliCredential(),
        ),
        instructions="You are a friendly assistant. Keep your answers brief.",
    )
    print(await agent.run("Which city is the capital of Karnataka?"))


if __name__ == "__main__":
    asyncio.run(main())

Choosing between Agent Framework and AG2

For teams already on Azure, .NET or Microsoft Foundry, Agent Framework is the natural choice because it is Microsoft's supported path and has first party .NET packages. A bank's IT team in Singapore or a government supplier in the UK that standardises on .NET will find it fits existing tooling, security reviews and procurement processes. Python teams on Azure benefit too, since both languages are covered by the same 1.0 commitment to stable APIs.

AG2 suits teams who liked AutoGen's conversational multi-agent style and want to stay with a community governed Python project. Because AG2 and Microsoft's projects have diverged, do not assume that code or documentation for one works with the other. Pin versions, read each project's changelog and plan a review whenever you upgrade, since both are moving quickly. AG2's own README describes the project as on the path to v1.0, so expect further changes.

Check your understanding