In 2026, picking an agent framework means parsing three names that sound interchangeable: Microsoft AutoGen (maintenance on GitHub), AG2 (community fork that ships the legacy autogen PyPI package), and Microsoft Agent Framework (the production successor Microsoft wants new projects on). The best agent framework for your team depends on orchestration style, cloud alignment, and whether you need conversation patterns or graph workflows.

Key takeaways:

  • Microsoft AutoGen (microsoft/autogen) is in maintenance mode: bug fixes and security patches only; new users are directed to Microsoft Agent Framework. (Source: AutoGen GitHub)
  • pip install autogen today resolves to AG2 (ag2==0.13.3), not Microsoft's autogen-agentchat 0.4 stack. Verify imports before you migrate. (Source: PyPI autogen)
  • Microsoft Agent Framework 1.0 (April 2026) is the vendor successor with stable graph workflows, MCP, and multi-agent patterns including Magentic-One. (Source: Microsoft Agent Framework 1.0)
  • Build 2026 (June 2-3) tied agents to enterprise context (Work IQ, Foundry IQ) and OS containment (MXC, OpenClaw on Windows). (Source: Build 2026 Live Blog)
  • LangGraph and CrewAI remain strong when you want provider-neutral graphs or fast role-based crews; provider SDKs win when you are already locked to one model vendor. (Inference: synthesis from ecosystem comparisons)

The point is not "which GitHub repo has more stars." It is which orchestration contract you can still debug at 2 a.m. in 2028 without a surprise package rename.

Primary sources: AutoGen GitHub README, AG2 GitHub, Microsoft Learn migration guide, Microsoft Agent Framework 1.0 announcement, Build 2026 live blog, PyPI autogen metadata (verified 2026-06-08).

What shipped

  • Maintenance mode on microsoft/autogen: The README carries an orange maintenance badge and a CAUTION block stating AutoGen receives no new features and is community managed; new users should start on Microsoft Agent Framework. (Source: AutoGen GitHub)
  • Microsoft Agent Framework 1.0 (April 8, 2026): GA for Python and .NET with stable APIs, graph-based Agent Workflows, MCP tool discovery, multi-agent orchestration (sequential, concurrent, handoff, group chat, Magentic-One), checkpointing, and published AutoGen and Semantic Kernel migration guides. (Source: Microsoft Agent Framework 1.0)
  • AG2 fork momentum: The ag2ai/ag2 repo reports roughly 4.6k GitHub stars, Apache 2.0 licensing for fork additions, and active releases (v0.13.3 on 2026-06-05). Founding authors Chi Wang and Qingyun Wu continue development under open governance at ag2.ai. (Source: AG2 GitHub)
  • PyPI name collision: The autogen PyPI package at v0.13.3 declares requires_dist: ag2==0.13.3, while Microsoft's 0.4 stack publishes separately as autogen-agentchat (v0.7.5 at verification time). (Source: PyPI autogen)
  • Build 2026 agent stack (June 2-3, 2026): Microsoft positioned Agent Framework 1.0 as GA, highlighted Microsoft IQ (Work IQ, Foundry IQ, Web IQ), previewed Microsoft Execution Containers (MXC) for policy-driven agent containment, and announced OpenClaw running on Windows through MXC. (Source: Build 2026 Live Blog)
  • Migration guide shipped: Microsoft Learn documents side-by-side AutoGen-to-Agent Framework mappings for model clients, single agents, GroupChat patterns, checkpointing, and observability. (Source: Microsoft Learn Migration Guide)

The three-way split: repo, PyPI, and successor

Practitioner payoff: most migration failures here are import errors, not architecture debates.

Path A: microsoft/autogen (maintenance). This repo hosts AutoGen 0.4's layered packages (autogen-core, autogen-agentchat, autogen-ext). Microsoft will accept bug fixes and security patches, but feature development moved to Agent Framework. If you already run AssistantAgent from autogen_agentchat, you are on this path. (Source: AutoGen GitHub)

Path B: AG2 on PyPI (ag2, autogen, pyautogen). The original research authors forked to ag2ai and kept releasing under the legacy PyPI names. That preserves conversable-agent and group-chat patterns familiar from pre-0.4 AutoGen, with Apache 2.0 governance outside Microsoft. (Source: AG2 GitHub)

Path C: Microsoft Agent Framework (agent-framework). This is Microsoft's unified successor, merging Semantic Kernel enterprise features with AutoGen orchestration ideas into typed graph workflows, hosted tools, and first-party Azure or Foundry connectors. (Source: Microsoft Learn Migration Guide)

Operator note (first-hand): On 2026-06-08, unauthenticated PyPI metadata showed autogen 0.13.3 with requires_dist pinning ag2==0.13.3, while autogen-agentchat resolved to 0.7.5 as a separate package. Running pip install autogen in a fresh venv therefore pulls AG2, not Microsoft's maintenance repo stack. Pin autogen-agentchat explicitly if you intend the Microsoft 0.4 line.

Microsoft Agent Framework: when the vendor path wins

Why this matters: Agent Framework is the only branch Microsoft commits to for long-term support and enterprise integration.

Decision rule: choose Microsoft Agent Framework when you are starting a new multi-agent system and can accept graph-first orchestration, especially on Azure, Microsoft Foundry, or .NET estates.

Key benefit: Version 1.0 locks stable workflow APIs, middleware hooks, declarative YAML agents, MCP integration, and checkpointed long-running processes. The April announcement positions A2A cross-runtime messaging with "A2A 1.0 support coming soon." (Source: Microsoft Agent Framework 1.0)

Practitioner payoff: The Learn migration guide maps AutoGen Team and GroupChat patterns to Agent Framework Workflow graphs. Expect a partial rewrite of orchestration code, not a semver bump. Model client setup stays familiar, but tool decorators, agent defaults, and multi-agent composition change shape. (Source: Microsoft Learn Migration Guide)

Defensive focus: Build 2026 adds operator context beyond the SDK. Work IQ and Foundry IQ supply organizational grounding; MXC supplies OS-level containment for agents like OpenClaw on Windows. Inference: teams already standardizing on Microsoft 365 plus Azure should treat MAF plus IQ plus MXC as one procurement story, not three optional add-ons. (Sources: Build 2026 Live Blog, Microsoft Agent Framework 1.0)

AG2: when the community fork wins

Why this matters: AG2 is the continuity bet for codebases that depended on classic AutoGen conversation loops and the original PyPI install path.

Decision rule: choose AG2 when you want open governance, Apache 2.0 licensing clarity, and minimal disruption to conversable-agent tutorials, but you can self-fund observability, hosting, and enterprise compliance.

Key benefit: AG2 ships active releases (v0.13.3 in early June 2026), supports MCP and multi-agent patterns, and documents itself as "Open-Source AgentOS." The fork's smaller star count versus microsoft/autogen reflects branding confusion, not necessarily engineering stagnation. (Source: AG2 GitHub)

Common mistake: assuming microsoft/autogen stars guarantee the package you installed. The star-rich repo and the PyPI alias point to different maintainers after the fork.

Defensive focus: AG2 does not ship Microsoft's Build 2026 enterprise envelope (Work IQ, Agent 365, MXC). Inference: AG2 fits research-heavy teams and startups with strong Python ops; regulated enterprises should budget extra for governance layers. (Inference: AgenticWire read)

LangGraph, CrewAI, and provider SDKs: the honest alternatives

The durable concept beneath the AutoGen fracture is framework fragmentation: no single repo owns "agents" anymore.

Framework | Orchestration style | Best for | Maintenance / vendor | Migration from AutoGen
Microsoft Agent Framework | Graph workflows + multi-agent | Azure/.NET teams, greenfield MAF | Microsoft GA 1.0 | Guided guide; partial rewrite
AG2 | Conversation / group chat | Legacy PyPI users, open governance | Community (ag2ai) | Lowest friction on ag2/autogen PyPI
autogen-agentchat (0.4) | Event-driven AgentChat | Existing 0.4 deployments | Microsoft maintenance | Stay put; plan MAF move
LangGraph | Directed graph / state machine | Production control, checkpointing | LangChain ecosystem | Redesign topology; no drop-in
CrewAI | Role-based crews | Fast multi-agent prototypes | CrewAI Inc. + OSS | Rewrite roles vs conversations
OpenAI Agents SDK | Handoffs + tool loop | OpenAI-first production | OpenAI vendor | New harness; see SDK coverage

LangGraph ranks highly in independent 2026 comparisons for production state machines and traceability; CrewAI wins when teams need a working crew in hours. Provider SDKs (OpenAI Agents SDK, Claude Agent SDK, Google ADK) tighten integration at the cost of neutrality. (Inference: synthesis from AlphaCorp developer guide and AgenticWire provider coverage)

Framework choice in 2026 is less about demo speed and more about who owns your orchestration contract when the original maintainers fork, rebrand, or enter maintenance.

Decision rules for teams picking in 2026

Greenfield, Microsoft-aligned: pip install agent-framework, follow the Learn quickstart, adopt MCP early for tool portability, and treat A2A as preview until 1.0 ships. (Source: Microsoft Agent Framework 1.0)

Greenfield, provider-neutral: prototype in CrewAI if roles fit; graduate to LangGraph when you need deterministic graphs, pause/resume, and audit trails. (Inference: AgenticWire read)

Legacy pip install autogen scripts: inspect import autogen provenance. If PyPI pulled ag2, stay on AG2 intentionally or pin autogen-agentchat and plan MAF migration. (Source: PyPI autogen)

Existing AutoGen 0.4 production: remain on maintenance branch for stability; schedule MAF migration per Learn guide; do not expect new AutoGen features. (Source: AutoGen GitHub)

Enterprise Windows agent deployments post-Build 2026: evaluate MXC containment for any autonomous runtime (OpenClaw, Copilot CLI, Claude Code) before approving wide install. (Source: Build 2026 Live Blog)

Context: how AutoGen fractured

AutoGen began as Microsoft Research work on multi-agent conversation, gained massive GitHub traction, and split along governance lines. Microsoft steered the flagship repo toward a 0.4 architecture and, by 2026, maintenance mode with Agent Framework as successor. The founding authors forked to AG2, retained PyPI package names, and argued open governance would speed community contribution. (Sources: AutoGen GitHub, AG2 GitHub)

That fracture mirrors a wider 2026 pattern: vendors productize agent harnesses and sandboxes while open-source frameworks compete on orchestration primitives. For how Microsoft packages graph workflows and MCP in the successor SDK, see our Microsoft Agent Framework 1.0 workflows and MCP coverage. For harness versus compute separation on long runs, see OpenAI Agents SDK harness versus sandbox.

FAQ

Is AutoGen still maintained in 2026?

Yes, but only in maintenance mode. Microsoft's README states AutoGen receives bug fixes and critical security patches, not new features, and directs new users to Microsoft Agent Framework. Community support continues on Discord and GitHub Discussions with variable response times. (Source: AutoGen GitHub)

What is AG2 and how is it different from Microsoft AutoGen?

AG2 is the community fork maintained by AutoGen's original authors under ag2ai. It continues development on the legacy PyPI packages (ag2, autogen, pyautogen). Microsoft's microsoft/autogen repo tracks the separate 0.4 autogen-agentchat architecture now in maintenance. (Sources: AG2 GitHub, AutoGen GitHub)

Should new projects start on Microsoft Agent Framework or AG2?

Microsoft Agent Framework if you want vendor-supported graph workflows, Azure or Foundry integration, and a published migration path from AutoGen. AG2 if you prioritize open governance, Apache 2.0 licensing, and continuity with classic conversable-agent APIs, and you will self-host governance. (Sources: Microsoft Agent Framework 1.0, AG2 GitHub)

Which agent framework is best for production in 2026?

There is no universal winner. LangGraph leads for explicit stateful graphs; Microsoft Agent Framework leads for Microsoft-centric enterprises; CrewAI leads for fast multi-agent prototypes; provider SDKs lead when you are already committed to one model vendor. Production readiness depends more on observability, checkpointing, and governance than GitHub stars. (Inference: AgenticWire read)

Can I keep using pip install autogen after the fork?

Yes, but you are installing AG2. PyPI's autogen 0.13.3 package depends on ag2==0.13.3. For Microsoft's 0.4 stack, install autogen-agentchat and related packages explicitly. (Source: PyPI autogen)

How do I migrate from AutoGen to Microsoft Agent Framework?

Follow Microsoft's Learn migration guide: map model clients first, then single agents, then multi-agent GroupChat patterns to Agent Framework workflows. Expect orchestration code changes where AutoGen Teams become typed workflows. Test checkpointing and observability early. (Source: Microsoft Learn Migration Guide)

Related coverage

References