Microsoft’s AI Frontiers Lab Is Over and AutoGen Gets No Fixes: How to Migrate (2026)

Microsoft’s AI Frontiers lab, the team behind the AutoGen agent framework, is over. Co-founder John Langford announced it on his blog on 6 November 2026, after the group was folded into Microsoft AI in September.

For developers, the bigger news sits in the AutoGen repo itself. A README change merged on 16 October says Microsoft is no longer providing bug fixes, security fixes, new features or support for AutoGen.

This guide covers what happened, which AutoGen package you are actually running, your three migration options, a step-by-step port to Microsoft Agent Framework with code, what does not work there yet, and how it compares with AG2 and the OpenAI and Claude agent SDKs.

What Happened to AI Frontiers

AI Frontiers (AIF) was a sub-organisation of Microsoft Research with about 50 people. Langford helped set it up in 2023, after ChatGPT arrived, to pull Microsoft’s research efforts together.

The group’s best-known output was AutoGen, one of the first popular multi-agent frameworks. It also built:

• Fara: a small computer-use agent model (MIT licence).
• Research work: the Next Latent training objective, the Dion3 optimiser (an improved Muon) and the Full Bandwidth Transformer architecture.
• Shared tooling: a model-building monorepo used across Microsoft Research.

In his post “The End of AI Frontiers” (6 Nov, updated 7 Nov), Langford writes that the group was reorganised into Microsoft AI in early September. He says that organisation’s leadership, which came from Inflection, focuses on scaling, and that “the core philosophy there is scaling, and aggressively not research.”

His conclusion is blunt: “So AI Frontiers is over.” He says he placed his team as best he could and left, and that there will be no AIF internships in the coming year.

Note: This is one founder’s account. We could not find a public Microsoft statement about AI Frontiers. Treat the reasons he gives as his view, not Microsoft’s.

Timeline: From AutoGen to “AI Frontiers Is Over”

What This Means for AutoGen

AutoGen did not die overnight. It had been winding down for a year: no release since September 2025, and the README already called it “maintenance mode” with a pointer to Agent Framework.

The October change went further. According to PR #8366 by Microsoft’s westey-m, the README now states that Microsoft is not providing bug fixes, security fixes, new features or support, and that the repo stays up for reference only. Later commits add that AutoGen was never meant for production use.

Some cached copies of the README still show the older “community managed” wording. Go by the merged PR.

Warning: The GitHub repo is not flagged as “archived” in GitHub’s own settings, so dependency scanners that check only that flag will not warn you. If AutoGen runs anything that touches the internet, user input or tools with real permissions, plan your exit now.

What still works

• Existing installs: your code keeps running; nothing is being pulled from PyPI.
• Licence: code stays MIT, docs CC-BY-4.0, so forking is allowed.
• Prototypes and teaching: fine for learning, as long as you accept unpatched bugs.

Check Which AutoGen You Are Running

Three different projects use the AutoGen name on PyPI. The warning only applies to Microsoft’s packages, so check first.

python -m pip list | grep -iE '^(pyautogen|autogen|ag2|agent-framework)'

On Windows PowerShell, use pip list | Select-String "autogen|ag2|agent-framework" instead.

Your Three Options

Step by Step: Port AutoGen to Agent Framework

Microsoft says Agent Framework is built by the same core AutoGen and Semantic Kernel teams. Its migration guide maps almost every AutoGen concept, but the defaults change, so test as you go.

1. Check Python. Agent Framework needs Python 3.11 or newer (3.11 to 3.15 supported). Create a fresh virtual environment so old AutoGen packages do not clash.
2. Install the framework. pip install agent-framework installs the core (OpenAI, Azure OpenAI, workflows) plus the standard integrations. For Foundry only, pip install agent-framework-foundry azure-identity is enough.
3. Set your key. The OpenAI client reads OPENAI_API_KEY from the environment. Agent Framework does not load .env files on its own, so add from dotenv import load_dotenv; load_dotenv() if you use one.
4. Swap the model client. Replace the AutoGen client import with from agent_framework.openai import OpenAIChatCompletionClient (or OpenAIChatClient for the Responses API).
5. Rename the agent. AssistantAgent becomes Agent, model_client= becomes client=, and system_message= becomes instructions=.
6. Convert tools. Replace FunctionTool(...) wrappers with the @tool decorator; schemas come from type hints.
7. Add sessions. Agents are stateless between calls. Create session = agent.create_session() and pass it to every run() that needs memory.
8. Rebuild teams as workflows. RoundRobinGroupChat maps to SequentialBuilder, GraphFlow to WorkflowBuilder, and MagenticOneGroupChat to MagenticBuilder.
9. Compare outputs and costs. Run the same prompts through both versions and count model calls (see the tool-loop note below).

Before and after: a single agent with a tool

AutoGen version:

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

client = OpenAIChatCompletionClient(model="gpt-5")
agent = AssistantAgent(
    name="assistant",
    model_client=client,
    system_message="You are a helpful assistant.",
    tools=[get_weather],
)
result = await agent.run(task="What's the weather in Delhi?")

Agent Framework version:

from typing import Annotated
from pydantic import Field
from agent_framework import Agent, tool
from agent_framework.openai import OpenAIChatCompletionClient

@tool
def get_weather(location: Annotated[str, Field(description="City name")]) -> str:
    """Get weather for a location."""
    return f"Weather in {location}: sunny"

client = OpenAIChatCompletionClient(model="gpt-5")
agent = Agent(
    name="assistant",
    client=client,
    instructions="You are a helpful assistant.",
    tools=[get_weather],
)
session = agent.create_session()
result = await agent.run("What's the weather in Delhi?", session=session)
print(result.text)

Before and after: a two-agent team

AutoGen’s round-robin chat becomes a sequential workflow. Note it is sequential, not a true round robin.

from agent_framework import SequentialBuilder

workflow = SequentialBuilder(participants=[writer, reviewer]).build()
async for event in workflow.run("Draft a launch email", stream=True):
    if event.type == "output":
        history = event.data  # list of Message objects

Tip: Agent Framework’s FileCheckpointStorage lets a workflow pause and resume after a crash or a human approval. AutoGen teams had nothing built in, so this is worth adopting during the port rather than after.

If you use Claude

Anthropic support is not in the stable package yet. A beta exists, updated weekly:

pip install --pre agent-framework-anthropic

PyPI marks it as a pre-release (1.0.0b261106, 7 Nov 2026) that “may not be stable for production use.” For production Claude agents, Anthropic’s own Claude Agent SDK is the safer choice today.

AutoGen to Agent Framework Cheat Sheet

Source: Microsoft Learn, “AutoGen to Microsoft Agent Framework Migration Guide”.

What Agent Framework Cannot Do Yet

Microsoft’s own guide lists these as missing or planned in the Python version. Check them against your AutoGen project before you commit to a date.

Note: If you run local models through Ollama, a temporary workaround is Ollama’s OpenAI-compatible endpoint with the OpenAI client. Microsoft does not document this as supported, so test tool calling carefully. Our best Ollama models list notes which models handle tools well.

Agent Framework vs AG2 vs Other Agent Frameworks

Versions for Agent Framework, AG2 and AutoGen are from PyPI on 9 November 2026. For the other two, see our OpenAI Agents SDK guide and Claude Agent SDK guide.

A note on AG2

AG2 grew out of the original AutoGen contributors and is now its own project under Apache 2.0. Install it with pip install "ag2[openai]" and add extras such as anthropic, gemini or ollama.

The 1.x ag2 package no longer ships the classic autogen import; that lives in a separate ag2-classic package. So AG2 is also a port, not a drop-in swap.

Migration Problems and Fixes

“My bill went up after porting”

AutoGen’s AssistantAgent made one tool round by default. Agent Framework’s Agent keeps calling tools until it has a final answer, so a single request can make several model calls. Log the call count and tighten tool descriptions or instructions.

“The agent forgot what I said one message ago”

You did not pass a session. Create one with agent.create_session() and pass session=session to each run().

“ModuleNotFoundError: agent_framework.anthropic”

The Anthropic client is a pre-release. Install it with pip install --pre agent-framework-anthropic; plain pip install skips pre-releases.

“My .env keys are ignored”

Agent Framework does not read .env files automatically. Call load_dotenv() from python-dotenv at the top of your script.

“Imports clash after installing both”

Old autogen-* packages and new ones can pull different OpenAI SDK versions. Port inside a clean virtual environment and only keep AutoGen in a separate one for side-by-side tests.

“My MCP server will not connect”

Agent Framework 1.23+ supports MCPStdioTool and MCPStreamableHTTPTool but not WebSocket transport. Switch the server to stdio or streamable HTTP; our Python MCP server tutorial covers both.

My Take

The lab drama will get the headlines, but the practical story is simpler: AutoGen is now unpatched software. Over 60,000 GitHub stars means a lot of tutorials, demos and quietly deployed bots still import it.

Agent Framework is a reasonable successor. It is clearly designed by people who knew AutoGen’s weak spots, adding sessions, checkpointing and typed workflows. Its weakness is model choice: shipping 1.25 releases without stable Claude or Ollama clients is an odd gap for a framework meant to replace a model-agnostic one.

If I were moving a small project this week, I would pick by model first. Azure or OpenAI: Agent Framework. Claude or local models: AG2, or a vendor SDK, and revisit Agent Framework when its clients go stable.

Langford’s wider point, that cutting research for pure scaling is a mistake, is a debate for another day. For developers, the lesson is older: check who maintains your agent framework before you build on it.

FAQ

Is Microsoft AutoGen discontinued?

Effectively yes. A README change merged on 16 October 2026 says Microsoft no longer provides bug fixes, security fixes, new features or support, and the last release was 0.7.5 in September 2025.

What happened to Microsoft’s AI Frontiers lab?

Co-founder John Langford wrote on 6 November 2026 that the ~50-person group was reorganised into Microsoft AI in September and “is over.” Microsoft has not commented publicly that we could find.

What replaces AutoGen?

Microsoft points users to Microsoft Agent Framework for Python and .NET. AG2 is a separate community framework that keeps the AutoGen-style approach.

Can I still use AutoGen?

Your code will keep running, and the MIT licence allows forks. But security issues may never be fixed, so Microsoft advises against production use.

Does Microsoft Agent Framework support Claude?

Only through a beta package, agent-framework-anthropic, installed with pip install --pre. The stable migration guide still lists Anthropic as planned.

Is AG2 the same as AutoGen?

No. AG2 started from AutoGen’s code and contributors but is an independent Apache 2.0 project with its own 1.x API.

How long does a migration take?

For one agent with a few tools, often an afternoon, since most changes are renames. Multi-agent teams take longer because group chats become typed workflows with different behaviour.

What happens to Fara?

Langford’s post does not say. The Fara repo is MIT-licensed and was last updated in September, so treat it like AutoGen: usable, but without guaranteed updates.

Sources

• John Langford, “The End of AI Frontiers”, hunch.net (6 Nov 2026, updated 7 Nov)
• GitHub microsoft/autogen, PR #8366 “docs: clarify AutoGen is archived and no longer maintained” (merged 16 Oct 2026) and README
• Microsoft Learn: AutoGen to Microsoft Agent Framework Migration Guide; Your First Agent quickstart
• Microsoft DevBlogs: Microsoft Agent Framework version 1.0 (3 Apr 2026)
• PyPI: agent-framework 1.25.0, agent-framework-anthropic 1.0.0b261106, autogen-agentchat 0.7.5, pyautogen, autogen, ag2 1.3.0
• DEV Community (Gabriel Anhaia): “The Microsoft lab that worked on AutoGen is over”
• AI/TLDR: “Microsoft Research closes AI Frontiers”
• Freakyjolly: Best MCP servers in 2026