AutoGen is Microsoft Research’s open-source framework for conversation-driven multi-agent systems, agents that talk through a problem instead of calling tools in sequence. Microsoft put it in maintenance mode on October 2, 2025. Three codebases now carry the lineage: legacy AutoGen, Microsoft Agent Framework, and AG2, a third-party fork rather than a Microsoft product. Orchestration frameworks define how an agent runs, not what its data means; Atlan’s context layer connects to all of them through MCP.
Most 2026 coverage of AutoGen is about leaving it. This page covers what still works and how to navigate the split.
| Field | Value |
|---|---|
| What it is | Open-source Microsoft Research framework for conversation-driven multi-agent systems |
| Created by | Qingyun Wu, Chi Wang, and Microsoft Research (arXiv, August 2023) |
| Current status | Maintenance mode since October 2, 2025, bug/security fixes only |
| Official successor | Microsoft Agent Framework (MAF), 1.0 GA April 3, 2026 (.NET and Python; Go in public preview) |
| Third-party fork | AG2, Apache 2.0, repo created November 11, 2024, administered by Chi Wang and Qingyun Wu |
| Current packages | autogen-agentchat, autogen-core, autogen-ext on the 0.4+ line; pyautogen (0.2) is no longer maintained by Microsoft |
| Best for | Research, prototyping, debate/critique workflows |
| GitHub stars | 61,042 (September 18, 2026), about 4.5x MAF’s 13,574 |
What is AutoGen?
AutoGen is Microsoft Research’s open-source framework for building multi-agent systems where agents reason together through structured conversation rather than a single model calling tools in a fixed sequence. The original paper, led by Qingyun Wu and Chi Wang (arXiv, August 2023), describes agents defined by role that converse to reach a solution, a framing Microsoft’s project page confirms. See what an agentic framework is for the baseline this cluster builds on.
The ConversableAgent pattern
Every AutoGen agent inherits from one primitive: ConversableAgent. Instances exchange messages, propose actions, and critique each other’s output, closer to colleagues working a problem together than a script executing steps.
User Proxy, Assistant, and Code Executor agents
Three named roles handle most AutoGen workflows: User Proxy represents the human, Assistant reasons through the problem and drafts solutions, Code Executor runs generated code and returns real output, errors included. Most deployments use only two or three of these roles.
GroupChat: how AutoGen’s agents talk to each other
For problems needing more than one specialist, AutoGen’s GroupChat pattern lets User Proxy, Assistant, and Code Executor agents share one conversation thread, with a manager routing turns between them. It carries the sharpest governance trade-off here: rich conversation, no built-in ledger of which agent’s contribution shaped the final answer.
AutoGen proved agents reasoning together produces different results than a controller calling tools in sequence. What it never specified is what those conversations are grounded in: the shared facts and lineage that keep three agents from confidently agreeing on three different numbers.
Is AutoGen deprecated? What maintenance mode means in 2026
AutoGen is not abandoned, but it’s no longer where new development happens. Microsoft announced maintenance mode on October 2, 2025, in Eric Zhu’s own words: the project “has a stable API and will continue to receive critical bug fixes and security patches.” The last feature release, python-v0.7.5, shipped two days earlier on September 30, 2025. The GitHub README carries the banner and now redirects newcomers to the Microsoft Agent Framework instead of AutoGen’s own quickstart.
One thing worth knowing before you check your own install: the banner lives on the GitHub README, not on microsoft.github.io/autogen. A reader who lands on the docs site gets no signal the project is frozen, and a “0.2 Docs” link is still there. If your imports are pyautogen or autogen-agentchat~=0.2, you are on the pre-0.4 line, which Microsoft describes as no longer maintained by Microsoft at all. Version 0.4, first stable on January 10, 2025, was a ground-up rewrite to an asynchronous, event-driven architecture, and its packages are autogen-agentchat, autogen-core and autogen-ext.
Despite the redirect, AutoGen’s star count tells a more nuanced story. The repository holds 61,042 GitHub stars as of September 18, 2026, about 4.5 times the Microsoft Agent Framework’s 13,574 read the same day. Legacy adoption doesn’t vanish overnight.
Maintenance mode answers whether AutoGen is dead, not whether moving off it fixes the underlying problem: neither its freeze nor its retirement addresses what agents built on it were allowed to know, a root cause behind why AI agents fail in production that survives the migration.
The AI Context Stack
A four-layer blueprint for the context infrastructure that sits underneath every agentic framework, AutoGen included, so a migration between frameworks never means rebuilding your business definitions from scratch.
Get the Context Stack BriefAutoGen, Microsoft Agent Framework, or AG2: which path should you take?
Since October 2025, “using AutoGen” means choosing between three paths: the original framework, Microsoft’s official successor, or the third-party fork. Most 2026 content mentions AG2 without explaining why the split happened, and a good deal of it calls AG2 a Microsoft rename. It is not one.
Star counts below were read on September 18, 2026.
| Path | Maintainer | License | Status | GitHub stars | Best for |
|---|---|---|---|---|---|
| Legacy AutoGen | Microsoft Research | MIT code, CC-BY-4.0 docs | Maintenance mode | 61,042 | Existing codebases |
| Microsoft Agent Framework | Microsoft | MIT | GA (April 3, 2026) | 13,574 | New enterprise builds |
| AG2 | Chi Wang, Qingyun Wu (third party) | Apache 2.0 | Active | 4,936 | Independent governance |
AG2’s roughly 4,900 stars, per its own repository, reflect credibility, not mass migration. Chi Wang called the motivation “the need to move faster and make decisions more efficiently without corporate constraints, and creating a more neutral space for contributions from various organizations”: project governance, not data governance.
Microsoft describes the Agent Framework as a separate successor SDK that absorbed concepts from both parents, “unifying the enterprise-ready foundations of Semantic Kernel with the innovative orchestrations of AutoGen into a single, open-source SDK”. It reached 1.0 GA on April 3, 2026 for .NET and Python, with Go in public preview. Semantic Kernel itself did not stop existing: the v1.x package still ships on its own repository under MIT.
Which path fits which team
Migrating to MAF suits new enterprise builds on Azure or a broader AI agent stack; see the deep dive for mechanics. Adopting AG2 suits teams wanting AutoGen’s model under independent governance. Staying on legacy AutoGen works for an existing codebase, provided the team accepts a repository that will ship no new features.
How does AutoGen compare to CrewAI and LangGraph?
AutoGen, CrewAI, and LangGraph each take a different approach to multi-agent system orchestration: conversation, role-based crews, explicit state graphs.
| Framework | Orchestration model | Context handling | Best fit |
|---|---|---|---|
| AutoGen | Conversation-driven (GroupChat, agent debate) | Conversation transcript as shared state | Research, debate and critique |
| CrewAI | Role-based crews | Unified Memory class, plus Memory Scopes and Slices across crews |
Structured business workflows |
| LangGraph | Explicit state graphs | Thread-scoped checkpointers and a cross-thread store | Complex, stateful pipelines |
Worth flagging: the context-handling row reflects Atlan’s own comparison content, not an independent benchmark, though it doesn’t mean AutoGen can’t connect to an external context layer, only that doing so takes more custom work than some peers.
None is categorically better. AutoGen fits debate and exploratory research; CrewAI’s role-based structure fits defined workflows with clear ownership; LangGraph’s state graphs fit complex pipelines needing checkpointing. Teams weighing all three can go deeper in the full agentic frameworks comparison or the OpenAI Agents SDK, LangChain, and CrewAI comparison.
Whichever model a team picks, what each agent may know still has to be governed somewhere, a gap covered in Atlan’s comparison of agent context layer tools.
When should you still use AutoGen in 2026?
Despite maintenance mode, AutoGen remains a legitimate, narrow choice, depending on what a team is building and how much production risk it can absorb.
| Criterion | Why it matters | What to look for |
|---|---|---|
| Use case (research vs. production) | Suits exploration, not scale | Debate, critique, research synthesis |
| Tolerance for maintenance-mode risk | No new features, on a frozen repository | Investment not ready to migrate off |
| Audit requirements | AutoGen’s docs describe no ledger of which agent’s contribution shaped an answer | Regulated use cases look elsewhere |
| MCP and context-layer integration effort | More wiring than peers with a documented MCP surface | Teams with in-house capacity |
| Migration cost if choosing legacy | Not gaining features | Budget for an eventual move |
Where it still fits, and where it doesn’t
Research and prototyping: the simplest pattern, one Assistant plus one User Proxy, carries most exploratory work before gaps compound at multi-agent scale. Debate and critique workflows are the design’s most natural fit: brainstorming, code review, iterative correction. When to choose differently: AutoGen Studio, the no-code UI, caps out per its original research paper at two-agent and GroupChat workflows. Otherwise, evaluate CrewAI, LangGraph, the Microsoft Agent Framework, or the selection guide.
Keeping AutoGen in 2026 is defensible, not a default: it fits debate-style work while the context question stays open, since the two are separate axes. AutoGen answers what an agent does next, not what it’s allowed to know, the gap behind AI agent accuracy failures, and neither does its successor or fork.
Context Gap Calculator
Score how much of your agent stack, AutoGen included, is reasoning over governed context versus definitions that live only in someone's head or a single framework's config.
Calculate Your Context GapAutoGen, Microsoft Agent Framework, AG2: three paths, one enterprise context gap
Whichever of the three paths a team chooses, the same question is left open: what agents may know, and how that’s governed. Atlan calls it the enterprise context gap.
GroupChat’s autonomy is AutoGen’s sharpest version of this gap. AutoGen’s documentation specifies the conversation mechanics, the manager that routes turns, and the termination conditions. It describes no ledger of which agent’s contribution shaped a final answer, no filter on what moves between agents, and no cost ceiling on a recursive exchange. Those controls are the implementer’s, a burden covered in Atlan’s guide to debugging multi-agent systems.
To be fair, transcripts let a reviewer roughly reconstruct one conversation without formal audit tooling, but that doesn’t scale to proving which data an agent was authorized to see across thousands of concurrent sessions. A readable log isn’t a governed one.
The strongest version of this argument isn’t that AutoGen is inferior to what replaced it. Look instead at where the successor’s governance work lands. Microsoft’s Agent Governance Toolkit, announced April 2, 2026 and currently in public preview, states its own boundary in architectural terms: it “enforces governance at the application middleware layer, not at the OS kernel level.” That is control over how an agent behaves. In Atlan’s reading, and this is our argument rather than Microsoft’s, it is a different question from what “ARR” means in a business or which table holds the canonical revenue figure. Atlan’s agentic frameworks comparison makes the same case across CrewAI and LangGraph.
If the designated successor still leaves this open, the gap isn’t a maintenance-mode problem, it’s what every agentic framework leaves ungoverned by design.
How Atlan approaches the enterprise context gap
GroupChat conversations and shared multi-agent memory work convincingly in a demo. In production, they create what Atlan calls multi-agent memory silos: no shared, governed source of truth, so teams rebuild business definitions every time they adopt a new framework or role.
Atlan’s context layer connects to AutoGen by name through its MCP server, alongside LangGraph and CrewAI, so the same certified definitions, lineage, and governance rules reach an agent regardless of which AutoGen path runs the orchestration. The Enterprise Data Graph is queryable by AutoGen agents at runtime, giving a User Proxy or Assistant agent the governed answer a human analyst would get. Teams can review the parallel context layer pattern built for CrewAI or the equivalent for LangGraph; the same architecture applies here.
One enterprise customer’s experience illustrates the pressure: it began a multi-agent build on AutoGen, then evolved its architecture as context requirements grew, consistent with teams “very conscious about not locking into anything including a single agentic framework,” as one put it. Portable context engineering makes that evolution possible without a full rebuild.
The framework decision is reversible: a team can move from AutoGen to AG2, MAF, or elsewhere within months. The context architecture underneath, what agents are allowed to know and who’s accountable for it, determines whether that migration is a rewrite or a redeploy.
AI Agent Context Readiness Checklist
See where an AutoGen, MAF, or AG2 deployment stands before the context gap becomes a production incident.
Check Your ReadinessWhat the fork and the successor leave to the context layer
AutoGen proved something real: agents reasoning together through conversation, not just calling tools in sequence, produces genuinely different results, an idea that shaped how the industry thinks about multi-agent design. Its own fork and its own official successor are both evidence of how far that idea traveled.
Neither the fork nor the successor closed the one gap that determines whether an agent works in production: what an agent is allowed to know, and who governs that knowledge as it changes. AutoGen, MAF, and AG2 orchestrate agents well. Orchestration defines how an agent runs, not what its data means, which makes the context layer a separate decision worth making deliberately rather than inheriting by default.
See how Atlan connects governed context to any agentic framework a team chooses, AutoGen included, and how to implement an enterprise context layer for AI end to end.
Sources
- AutoGen: Multi-Agent Conversation Framework, arXiv
- AutoGen, Microsoft Research
- microsoft/autogen repository, GitHub
- AutoGen is now in maintenance mode, Discussion #7066, GitHub
- microsoft/autogen releases, GitHub
- AutoGen v0.2 to v0.4 migration guide, Microsoft
- AutoGen .NET implementation, GitHub
- microsoft/agent-framework repository, GitHub
- Microsoft Agent Framework version 1.0, Microsoft Agent Framework devblog
- Semantic Kernel and Microsoft Agent Framework, Microsoft Agent Framework devblog
- microsoft/agent-governance-toolkit repository, GitHub
- ag2ai/ag2 repository, GitHub
- AutoGen Studio: A No-Code Multi-Agent Tool, arXiv
- microsoft/autogen LICENSE (CC-BY-4.0), GitHub
- microsoft/autogen LICENSE-CODE (MIT), GitHub
FAQs about AutoGen
1. What is the purpose of AutoGen?
AutoGen’s purpose is to let multiple AI agents collaborate through conversation, rather than one agent calling tools in sequence, for code generation and research synthesis. It’s built around the ConversableAgent primitive, where a User Proxy, an Assistant, and a Code Executor exchange messages until resolved.
2. Is AutoGen completely free to use?
Yes. AutoGen’s code is MIT, granted in the repository’s LICENSE-CODE file, while the repo-root LICENSE is Creative Commons BY 4.0 and covers the documentation. That split is why GitHub reports the repository licence as CC-BY-4.0. It is free to download, run, and modify. Microsoft doesn’t charge for the framework; costs come from the compute and API usage it consumes.
3. What is the difference between AutoGen and AutoGPT?
AutoGen is Microsoft Research’s structured multi-agent framework with defined agent roles (User Proxy, Assistant, Code Executor). AutoGPT is an earlier, more experimental single-agent project with less structured coordination. AutoGen has a stronger academic pedigree and clearer production patterns.
4. What is the AG2 fork, and should you use it instead of AutoGen?
AG2 is a third-party fork of AutoGen, not a Microsoft product and not a rename. Its repository was created on November 11, 2024 under Apache 2.0 and is administered by original creators Chi Wang and Qingyun Wu; microsoft/autogen still ships separately. Use AG2 for AutoGen’s model under independent governance. Stay on legacy AutoGen only for codebases not yet ready to migrate.
5. Who created AutoGen?
AutoGen was created by Microsoft Research, with Qingyun Wu and Chi Wang as lead authors, published as an arXiv paper in August 2023. Microsoft Research’s own project page carries the same attribution.
6. Does AutoGen support languages other than Python?
AutoGen is primarily a Python framework, though a .NET implementation ships in the same repository under Microsoft.AutoGen.*, with APIs Microsoft describes as not yet stable. Microsoft Agent Framework reached 1.0 for .NET and Python, with Go in public preview.
7. Does Atlan integrate with AutoGen?
Yes. Atlan’s context layer connects to AutoGen by name through its MCP server, exposing the same governed definitions and lineage LangGraph and CrewAI agents draw from, so AutoGen agents reason over current, audited context instead of duplicated ones.
