AutoGen v0.4 vs. AG2: The Microsoft Community Split Explained

AutoGen vs. AG2 refers to the split of the original AutoGen project into two lines: Microsoft rebuilt AutoGen with version v0.4 on an event-driven, actor-based architecture, while a community fork continued the classic, conversation-based code under the name AG2. Both lines share the basic concept of conversation-based multi-agent systems, but differ in origin, governance, architecture and roadmap. For DACH decision-makers, what counts most is which path is future-proof for their own project.

• AutoGen v0.4 is Microsoft's re-architecture: event-driven, actor-based, designed for scaling and distributed agents.
• AG2 is the community fork whose AG2 Classic line continues the v0.2 API.
• Strategically, Microsoft consolidates its agent roadmap in the Microsoft Agent Framework 1.0 (GA since April 3, 2026, MIT license), which brings together Semantic Kernel and AutoGen. AutoGen itself is in maintenance mode.

Origin and Governance: How the Split Came About

AutoGen originally emerged as a research and open-source project from the Microsoft environment and quickly became one of the best-known frameworks for multi-agent conversations. The original line (around version v0.2) popularized the framework through so-called conversable agents: agents that "talk" to each other in natural language, negotiate tasks and call tools.

With the further development to v0.4, Microsoft opted for a fundamental architectural rebuild rather than an evolutionary continuation. This break (a new architecture and a changed API philosophy) was the trigger for the community split. In November 2024, AG2 diverged from the AutoGen codebase to continue the conversation-centric line. The AG2 repository names Chi Wang and Qingyun Wu as its administrators, both authors of the original AutoGen paper. This created two parallel lines:

• AutoGen (Microsoft): repository at Microsoft, strategically embedded in the company's larger agent portfolio. Since entering maintenance mode, it has been maintained by the community.
• AG2 (ag2ai): independent project with its own roadmap. Its classic line, AG2 Classic, continues the v0.2 API for existing users.

Important for context: Microsoft consolidated its agent strategy in 2026. The Microsoft Agent Framework 1.0 has been generally available since April 3, 2026 (public preview since October 1, 2025) and unites Semantic Kernel and AutoGen under the MIT license. Before the merger, the predecessor projects together accumulated over 75,000 GitHub stars. AutoGen concepts are integrated into the new graph workflow of the Agent Framework. According to the AutoGen repository, AutoGen is in maintenance mode: no new features, and the community maintains it going forward. There, Microsoft points new users to the Agent Framework and existing users to a migration guide.

Architecture Differences: Conversation vs. Actors

The technical core of the split lies in the architecture model.

AutoGen v0.4 pursues an event-driven, actor-based approach. Instead of a linear conversation flow, agents communicate via messages and events. This resembles classic actor models from distributed programming: each agent is a self-contained unit that reacts to incoming messages. The advantages are better scalability, asynchronicity and suitability for distributed systems. The price is higher conceptual complexity compared to the simple "two agents talk to each other" model.

AG2 Classic sticks to the conversation-centric model of the original line. The central building blocks are conversable agents that collaborate in group chats and via handoffs. This model is more intuitive for beginners and sufficient for many prototyping scenarios: the mental load is lower, and the API is more familiar to anyone who has already worked with classic AutoGen.

This distinction is more than cosmetic. It determines how you structure multi-agent logic, how you debug and how well the system scales under load.

Table: AutoGen v0.4 vs. AG2 Compared

Note: Specific version states and star counts of both projects are snapshots and should be checked directly in the respective repository before any production decisions (as of 2026).

Conversation Agents: What Stays, What Changes

Both lines descend from the same conceptual root: agents that coordinate via natural language, call tools and divide tasks among several roles. In AG2 Classic this pattern is the main purpose, and an assistant agent, a user proxy agent and a group chat manager are enough to set up a collaborative pipeline.

In AutoGen v0.4 the same functional goal exists, but is realized via the event layer: conversations are a pattern on top of the message backbone, not the foundation itself. In the Microsoft Agent Framework, this line ultimately merges into the graph workflow, which models orchestration patterns (sequential, concurrent, handoff, group chat, Magentic) explicitly as a graph, i.e. more structured and closer to production than free-form conversation.

Concrete Example: Decision for a DACH SME Project

A Vienna-based B2B software company wants to build a multi-agent process for automated proposal generation: a research agent, a calculation agent, a review agent. The team faces the choice.

• Existing PoC in classic AutoGen (v0.2): The existing prototype runs with conversable agents. If it is to live on without an architecture break, AG2 Classic is the pragmatic path, since the API is almost identical and the migration effort is minimal.
• Strategic rebuild within the Microsoft stack: If the company already uses Azure and needs LTS guarantees, native MCP and A2A support as well as hosting in Europe, the Microsoft Agent Framework 1.0 is the more reliable choice. It ships connectors for seven providers (Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic, Bedrock, Gemini, Ollama); if you use Azure, you choose European regions (including Germany West Central in Frankfurt, West Europe, Sweden Central, Switzerland North).
• Strict sovereignty: What matters is where models and data run. Self-hosting with local models works with AG2 as well as with the Microsoft Agent Framework via Ollama. If you also want a roadmap that does not depend on Microsoft, choose AG2 or another self-hostable open-source framework.

Rule of thumb: existing v0.2 workloads to AG2 Classic (autogen package). Strategic rebuild within the Microsoft/Azure stack to the Microsoft Agent Framework. Pure AutoGen v0.4 is today mainly the bridge for existing Microsoft AutoGen customers toward the Agent Framework.

Migration: Realistic Paths

The migration path recommended by Microsoft is AutoGen to the Microsoft Agent Framework (AutoGen itself is in maintenance mode, and Microsoft provides a migration guide). Since AutoGen concepts are absorbed into the Agent Framework's graph workflow, the migration concerns the agents' business logic less than the remodeling of the orchestration as well as tool and provider wrapping.

Anyone who, on the other hand, wants to stay on the classic, conversation-based line does not need a forced migration. Since AG2 1.0 (released July 27, 2026), this line has been called AG2 Classic: it keeps the autogen import and classes such as ConversableAgent and GroupChat. According to the AG2 repository, AG2 1.0 itself is a newly built framework imported as ag2 and not a drop-in upgrade: the agent model, orchestration and imports have changed.

The practical trap lies in the package names on PyPI. Up to version 0.14, the ag2 package shipped the autogen import. From version 1.0 it no longer does, so an update without a version pin pulls in 1.0 and breaks existing classic code. Anyone who installed classic code via ag2 should pin ag2<1.0 or switch to the autogen package. AG2 manages that package as the classic codebase, which only receives security and critical bug fixes. pyautogen belongs to Microsoft again and today installs the current autogen-agentchat line. Anyone still running v0.2 code with it has to pin pyautogen~=0.2.0. Microsoft's v0.4 line runs under autogen-agentchat, autogen-core and autogen-ext.

For Agencies and B2B Decision-Makers

The lesson from the AutoGen/AG2 split for marketing agencies and DACH B2B teams: choose based on stability, talent pool and hosting, not on GitHub stars. Anyone who has to operate client projects over the long term benefits from the LTS commitment of the Microsoft Agent Framework. It can be hosted in European Azure regions or fully self-hosted, including with local models via Ollama. The CLOUD Act debate concerns the cloud and model services of US providers you use, not the SDK. Anyone who wants to keep an existing v0.2 system running without disruption is well served by AG2 Classic, but has to plan for the fact that it only receives security and critical bug fixes. Anyone building new projects independently of Microsoft's roadmap should evaluate AG2 1.0 or another self-hostable open-source framework.

As an agency in Vienna, Blck Alpaca supports framework evaluation, GDPR-compliant architecture and the migration of existing AutoGen systems, from the decision matrix to the production multi-agent pipeline. Talk to us before you commit technically: setting the right course at the start saves the expensive re-platforming effort later on.