Head to head
AutoGen vs Microsoft Agent Framework (2026)
Already running AutoGen in production and it's doing its job, with no near-term need for the newer capabilities → keep it; it still receives security fixes, and Microsoft's own migration guide means you aren't locked in if that changes.
Both checked against their own sites 19 September 2026 · October AI answer share: AutoGen 7%, Microsoft Agent Framework 3%
Reviewed by The Agents Index editorial team.
How AutoGen and Microsoft Agent Framework compare
| Question | ||
|---|---|---|
| The basics | ||
| Price | AutoGen: Freegithub.com | Microsoft Agent Framework: Freegithub.com |
| Category | AutoGen: Agent frameworks | Microsoft Agent Framework: Agent frameworks |
| Pricing model | AutoGen: free | Microsoft Agent Framework: freemium |
| Model / LLM | AutoGen: Model-agnostic (OpenAI/Azure first-class) | Microsoft Agent Framework: Model-agnostic (Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, Ollama, Foundry Local) |
| Bring your own model | AutoGen: Not stated | Microsoft Agent Framework: Yes |
| Interface | AutoGen: Library | Microsoft Agent Framework: Library |
| Deployment | AutoGen: Self-host | Microsoft Agent Framework: Both |
| Tags | AutoGen: API, Autonomous, Claude, Free tier, GPT, Multi‑agent, No‑code, Open models, Open source, Self‑hosted | Microsoft Agent Framework: API, Autonomous, Claude, Free tier, GPT, Multi‑agent, Open models, Open source, Self‑hosted |
| What it does | ||
| Writes code | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Debugs and fixes | AutoGen: Partlygithub.com | Microsoft Agent Framework: Not stated |
| Reviews code | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Not stated |
| Whole-codebase work | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Ships it | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Done-for-you service | AutoGen: Nogithub.com | Microsoft Agent Framework: Nolearn.microsoft.com |
| How much it does unattended | ||
| Runs autonomously | AutoGen: Yesgithub.com | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Multi-agent | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Agent permissions | AutoGen: Partlygithub.com | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Which models it runs on | ||
| Claude | AutoGen: PartlyCurrently limitedmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| GPT | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Open models | AutoGen: PartlyCurrently limitedmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Model choice | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Where you use it | ||
| In your editor | AutoGen: Nogithub.com | Microsoft Agent Framework: Nolearn.microsoft.com |
| On the command line | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| In a browser | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: PartlyCurrently limitedlearn.microsoft.com |
| Whose machine it runs on | ||
| Self-hosted | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| Open source | AutoGen: Yesgithub.com | Microsoft Agent Framework: Yesgithub.com |
| What it costs to run | ||
| How it meters | AutoGen: Partlygithub.com | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Free tier | AutoGen: Yesgithub.com | Microsoft Agent Framework: Yesgithub.com |
| API access | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
| MCP server | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Bring your own key | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yesgithub.com |
| Buying it for a team | ||
| Admin controls | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Audit log | AutoGen: Partlymicrosoft.github.io | Microsoft Agent Framework: Partlylearn.microsoft.com |
| Single sign-on | AutoGen: Nomicrosoft.github.io | Microsoft Agent Framework: Not stated |
| What happens to your code | ||
| Getting out | AutoGen: Yesmicrosoft.github.io | Microsoft Agent Framework: Yeslearn.microsoft.com |
Microsoft AutoGen popularized the "agents talking to agents" pattern and is still one of the most-starred agent frameworks on GitHub, but Microsoft itself has moved on: AutoGen has been in maintenance mode (bug and security fixes only) since late 2025, and its official successor, Microsoft Agent Framework, reached a production-ready 1.0 GA release on 2026-04-03. Microsoft didn't just add a new option alongside AutoGen: it put AutoGen AND its other agent stack, Semantic Kernel, into maintenance mode specifically to unify new development into Agent Framework. That makes this less a "pick a side" comparison than most framework match-ups on this site, and more a migration-timing decision for anyone with an existing AutoGen investment.
The GitHub activity, checked live on 2026-07-27, tells the same story. AutoGen's root repository has 60,020 stars and 9,032 forks, but its most recent commit to the main branch (2026-04-06) only updated the README's maintenance banner, and its latest tagged release remains python-v0.7.5, published 2025-09-30, unchanged for ten months. Agent Framework, by contrast, has 12,432 stars and 2,087 forks and is shipping today: its repository was pushed to as recently as this morning, and it tagged a new release (python-github-copilot-1.0.0) on 2026-07-23 for its GitHub Copilot SDK integration. There is also a third path worth knowing about: AG2, a fork maintained by AutoGen's original authors, controls the old PyPI packages, is licensed Apache-2.0, and had its own commit pushed today, genuinely active, not a dormant mirror. Anyone with AutoGen code today actually has two live forward paths, not one.
The technical gap is bigger than a rename. Agent Framework unifies AutoGen's multi-agent orchestration with Semantic Kernel's enterprise plumbing, and ships production-operations features AutoGen never had: an Agent Harness for guarded shell/filesystem access and human-in-the-loop approval, automatic context/token compaction, and an optional Foundry Hosted Agents layer for managed, scale-to-zero deployment with built-in OpenTelemetry observability. It also adds a CodeAct mode (the model writes a short program to batch tool calls instead of issuing them one at a time) that Microsoft's own BUILD 2026 benchmarks put at 52.4% faster completion and 63.9% fewer tokens on multi-step workloads, plus native support for both the Model Context Protocol and the Agent2Agent protocol. Model support broadens too: AutoGen is first-class for OpenAI and Azure OpenAI, while Agent Framework adds first-class Anthropic Claude and local Ollama models, plus Google Gemini through its Go SDK, and the language surface grows from Python and .NET to Python, .NET and Go. Neither framework charges a license fee: both are free to self-host and MIT-licensed at the core, so migrating costs engineering time, not a new subscription. Agent Framework's only paid layer, Foundry Hosted Agents, is optional and billed as ordinary Azure/Foundry consumption, the same "free core, pay your own way" shape AutoGen already has.
Who each one is for
AutoGen
Python or .NET developers and researchers exploring multi-agent orchestration who want a mature, free, self-hosted framework.
From Free
Watch for: It is a code library, not a hosted service. You build, host and operate the agents yourself.
Microsoft Agent Framework
.NET or Python teams already invested in Azure or Microsoft Foundry who want one supported agent SDK, an optional managed hosting path, and a documented migration off AutoGen or Semantic Kernel.
From Free
Watch for: It is a code library plus an optional managed layer, not a no-code product. You (or Foundry Hosted Agents) run and operate it.
Verdict: AutoGen or Microsoft Agent Framework?
Already running AutoGen in production and it's doing its job, with no near-term need for the newer capabilities → keep it; it still receives security fixes, and Microsoft's own migration guide means you aren't locked in if that changes. Actively extending an AutoGen codebase, or want to keep its exact API without a rewrite → look at the AG2 fork first: it's shipping real commits weekly and controls the original PyPI packages. Starting any new multi-agent project, or need production-ops features AutoGen never shipped (guarded shell/filesystem access, human-in-the-loop approval, managed hosting, native MCP/A2A support) → Microsoft Agent Framework; it's the stack Microsoft is now putting new engineering hours into. Running a large, mature AutoGen deployment with no urgent trigger to move → plan a deliberate migration using Microsoft's own from-AutoGen guide rather than a rewrite. Frozen is not the same as abandoned, and there's no cost pressure forcing an immediate switch.
Frequently asked questions
- Is Microsoft AutoGen dead?
- Not dead, but frozen: it's in maintenance mode, receiving only bug and security fixes. Its last tagged release is python-v0.7.5 from 2025-09-30, and its most recent commit (2026-04-06) only updated the README to redirect new users to the Microsoft Agent Framework. It still runs and is still supported for existing deployments: it just isn't where new features land.
- What is the actual difference between AutoGen, Agent Framework and Semantic Kernel?
- AutoGen and Semantic Kernel were two separate Microsoft agent stacks: AutoGen for multi-agent conversation patterns, Semantic Kernel for enterprise integration plumbing. Microsoft Agent Framework unifies both into one supported SDK for .NET, Python and Go, and has put both predecessors into maintenance mode to concentrate new development in one place. It reached a production-ready 1.0 GA release on 2026-04-03.
- Can I keep using AutoGen's exact API if I don't want to migrate yet?
- Yes. AG2, a community fork maintained by AutoGen's original authors, keeps the familiar API, controls the old PyPI packages, is licensed Apache-2.0, and is genuinely active (commits as recent as today). It's a real third option alongside "stay on frozen AutoGen" or "migrate to Agent Framework."
- Is Microsoft Agent Framework actually free?
- Yes. The SDK is MIT-licensed and free to self-host: you pay only your model provider. The only paid layer is the optional Foundry Hosted Agents managed deployment, billed as ordinary Azure/Microsoft Foundry consumption rather than a separate flat subscription, and it's not required to use the framework.
- Does Microsoft provide an official path for migrating off AutoGen?
- Yes. A first-party migration guide (learn.microsoft.com/en-us/agent-framework/migration-guide/from-autogen) walks existing AutoGen projects through moving onto Agent Framework, covering the conceptual mapping between AutoGen's agents/teams and Agent Framework's equivalents.
Sources and method
Every price here was read on the maker's own pricing page.
- AutoGen: github.com/microsoft/autogen/blob/main/LICENSE-CODE, checked 19 September 2026
- Microsoft Agent Framework: github.com/microsoft/agent-framework/blob/main/LICENSE, checked 19 September 2026
At a glance
AutoGen pioneered Microsoft's multi-agent conversation pattern but has been in maintenance mode since late 2025, while Microsoft Agent Framework is its official, actively-shipping successor, so the real question for most teams isn't which is better, it's when and how to migrate.
What changed in the index
Agents we’ve just listed and every plan whose price changed.