AG2 v1.0 Split: AG2 Classic Becomes a Fourth AutoGen Lineage

What changed in AG2 v1.0

AG2, the community fork of AutoGen run by the framework's original authors, has reorganized itself around a new architecture with its v1.0 release, and the change goes beyond a version bump. The top-level ag2 package is now a protocol-driven framework imported as ag2. It no longer ships the familiar autogen import name or the classic agent classes that defined the project from its start 2. Developers looking for ConversableAgent, GroupChat, or import autogen are directed to a separate distribution, AG2 Classic, installed with pip install ag2-classic 2.

The project is direct about compatibility. It calls v1.0 "not a drop-in upgrade" from Classic, because the agent model, orchestration approach, and imports have all changed. It points users to a group chat migration guide 2. AG2 Classic keeps the full legacy surface: the autogen.* namespace, AssistantAgent, UserProxyAgent, GroupChatManager, swarms, register_function, LLMConfig/OAI_CONFIG_LIST, and nested and sequential chat patterns 4. The current PyPI release of ag2 is 1.0.3 4. One report places the v1.0.0 release in July 2026 5.

How we got to four lineages

The split matters because of how many branches AutoGen already had. By late 2024 the original AutoGen line had forked. Chi Wang and Qingyun Wu, the primary authors of the AutoGen paper, set up AG2 as an independent, openly governed spin-off at ag2ai/ag2 1. Its licensing has two layers: new modifications are Apache 2.0, while inherited AutoGen code stays under MIT 1. On PyPI, the project lists the two as administrators and describes maintenance by volunteers from several organizations 4.

Microsoft took a different route. It merged AutoGen's research-oriented multi-agent ideas with Semantic Kernel, its enterprise SDK, into the Microsoft Agent Framework. Semantic Kernel launched in March 2023 with a C#-first design, added agent abstractions during 2024, and reached general availability for its Agents in April 2025 with sequential, concurrent, group-chat, handoff, and Magentic-style orchestration 1. Microsoft's own AutoGen has been in maintenance mode since October 2, 2025, with no new releases planned 5.

That left three lineages: Microsoft's frozen AutoGen, Microsoft's successor Agent Framework, and AG2 5. AG2 v1.0 effectively adds a fourth, because the fork now has its own legacy branch (Classic) next to its new protocol-driven core.

The naming problem gets worse

Package names are where this gets confusing. AG2 controls the autogen name on PyPI 5. That package, now at 0.14.1, describes itself as "the classic AutoGen repository," built around ConversableAgent, in maintenance mode with only security and critical bug fixes, and points users to ag2 as the actively developed version 3. The ag2 README, meanwhile, says Classic "remains maintained and installable" and that existing code "keeps working" 2.

Those statements don't strictly contradict each other, but they suggest different things. "Maintained" in the AG2 repo reads as reassurance. "Security and critical bug fixes only, with no new features" on the autogen package page reads as a wind-down 3. It is also not entirely clear from the public listings how the autogen and ag2-classic distributions relate to each other. A developer could reasonably end up with three package names (autogen, ag2-classic, ag2) and two quite different programming models, all under AG2's stewardship.

Why it matters for teams

Migration off AutoGen was already expensive. Moving to Microsoft's Agent Framework means rebuilding conversation-centric GroupChat designs as graph-based workflows, which one analysis calls "a conceptual rewrite rather than an import swap" 5. Staying with AG2 used to look like the low-friction option, since the classes and imports carried over. That is no longer true at v1.0. Teams on AG2 now face a similar choice: pin Classic and accept a maintenance-mode future, or rewrite for the new protocol-driven core 23.

Our read

AG2's reasoning is easy to follow. Separating a new architecture from an inherited one is cleaner than carrying the legacy API indefinitely, and putting the warning at the top of the README and PyPI page is responsible 24. The result, though, undercuts the main selling point the fork had against Microsoft's consolidation pitch, which was continuity. With Microsoft's original frozen, its successor requiring a redesign, and AG2 now running both a frozen-ish Classic and a non-compatible v1.0, no part of the AutoGen family offers a painless path forward.

For anyone with production AutoGen code, the practical advice is to pin versions explicitly, treat any move as a rewrite, and pick a destination based on which programming model is likely to be supported longest, not on which package still has "autogen" in its name.