AutoGen and Microsoft Agent Framework

Microsoft's conversation-first multi-agent framework AutoGen (now archived) and its successor, Microsoft Agent Framework, which merges AutoGen with Semantic Kernel around agents and typed graph workflows.

Contents

AutoGen is Microsoft Research’s open-source framework for building LLM applications from multiple conversable agents that talk to each other, to humans and to tools to solve a task (Wu et al., 2023). It popularised group-chat orchestration and code-writing/code-executing agent pairs. As of 2026-11 AutoGen is archived: it gets no bug fixes, security fixes or support, and Microsoft advises against production use (README). Its successor is Microsoft Agent Framework (MAF), built by the AutoGen and Semantic Kernel teams. MAF combines AutoGen’s simple agent abstractions with Semantic Kernel’s enterprise features and adds typed graph workflows (MAF overview). MAF 1.0 went GA for .NET and Python on 2026-04-03 (MAF 1.0). This page covers both, because you will meet AutoGen code in existing projects and should build new work on MAF.

Lineage

flowchart LR
  P["AutoGen v0.2 (2023): ConversableAgent, GroupChat"] --> V4["AutoGen v0.4 (Jan 2025): async event-driven Core + AgentChat + Extensions"]
  V4 --> ARC["AutoGen archived (reference only)"]
  SK["Semantic Kernel"] --> MAF["Microsoft Agent Framework 1.0 (Apr 2026): Agents + Workflows + Harness"]
  V4 --> MAF

AutoGen’s architecture (v0.4 line)

v0.4 was a ground-up rewrite. v0.2 users had complained about architectural limits, a messy API and weak debugging (v0.4 blog). It has three layers:

LayerWhat it is
Core (autogen-core)Asynchronous, event-driven message passing between agents (RoutedAgent + message handlers), with local and experimental distributed runtimes, Python and .NET
AgentChat (autogen-agentchat)High-level API: AssistantAgent, UserProxyAgent, teams such as RoundRobinGroupChat, SelectorGroupChat (an LLM picks the next speaker), Swarm (handoffs), MagenticOneGroupChat, experimental GraphFlow, plus termination conditions
Extensions (autogen-ext)Model clients (OpenAI, Azure, Anthropic, Ollama), code executors (Docker, Azure), MCP

Sources for the table: README, v0.4 blog, migration guide (class names), Notebook (SelectorGroupChat).

It also ships AutoGen Studio (a low-code GUI, explicitly not production-ready) and Magentic-One, a generalist multi-agent system for web and file tasks (README, v0.4 blog).

The coordination style is emergent. Agents post to a shared conversation, a selector or round-robin policy picks who speaks next, and a termination condition (such as TextMentionTermination("TERMINATE") or StopAfterNMessages) ends the run. This makes AutoGen flexible for code-heavy and novel topologies, but you must supply firm stop conditions (Notebook).

Microsoft Agent Framework

MAF has four areas (MAF overview):

  • Agents. An Agent wraps a chat client with instructions and tools (@tool, hosted tools, MCP). Unlike AutoGen’s AssistantAgent, which is single-turn unless max_tool_iterations is raised, an MAF Agent keeps calling tools until it can answer (migration guide). AgentSession holds conversation state, context providers supply memory, and middleware intercepts agent actions.
  • Harness agent. An opinionated long-task agent with planning and todos, context compaction, file access and memory, and tool approval.
  • Workflows. Explicit execution paths. Graph workflows (WorkflowBuilder) connect typed executors with edges (conditional, fan-out, fan-in), emit events, and checkpoint at superstep boundaries. Python also has an experimental functional @workflow API that uses native control flow (workflows).
  • Integrations. Model providers, memory stores, evaluation and UI adapters.

Built-in multi-agent orchestrations: sequential, concurrent, handoff, group chat and Magentic (a manager agent dynamically coordinates specialists). They support human-in-the-loop through approval-required tools and request-info pauses (workflows).

What changed from AutoGen

AspectAutoGenAgent Framework
Programming modelDual: low-level Core or high-level TeamSingle Workflow abstraction (graph or functional)
Graph semanticsGraphFlow: control flow, messages broadcast, edges are transitionsData flow: typed messages routed on edges activate executors
Node typesAgents onlyAgents, functions, sub-workflows
Human-in-the-loopNo built-in pause in TeamsRequest/response pauses (ctx.request_info, RequestInfoExecutor)
CheckpointingNone built inSuperstep checkpoints, resume
RuntimeEmbedded + experimental distributedSingle-process today; distributed planned
Agent loopSingle-turn by defaultMulti-turn tool loop by default

Source: migration guide. The mapping: RoundRobinGroupChat becomes SequentialBuilder, and MagenticOneGroupChat becomes MagenticBuilder (with max_round_count and stall limits).

Contradiction: The AutoGen migration guide lists Anthropic and Ollama clients as “planned” and Swarm-style handoff and SelectorGroupChat as “future patterns” (migration guide). The 1.0 announcement and the overview list Anthropic Claude and Ollama connectors and a handoff orchestration as available (MAF 1.0, workflows). The guide is probably stale in places, so check the current docs.

Status at 1.0 (2026-04)

Stable: agent abstraction and connectors (Foundry, Azure OpenAI, OpenAI, Anthropic, Bedrock, Gemini, Ollama), middleware, memory and context providers (Mem0, Redis, Neo4j and more), graph workflows with checkpointing, the five orchestrations, declarative YAML agents and workflows, and MCP. A2A 1.0 was announced as “coming soon”. Preview: DevUI, Foundry hosted agents, AG-UI, skills, the Agent Harness, and GitHub Copilot SDK and Claude Code SDK harness connectors (MAF 1.0). The Go SDK is in public preview (MAF overview).

Limitations and failure modes

  • AutoGen is unmaintained. Unpatched vulnerabilities may remain. Migrate before production (README).
  • Conversation-driven orchestration wanders. Group chats can loop or let agents talk past each other without strict termination, which costs tokens and latency (see cost and latency optimization) (Notebook).
  • Migration isn’t mechanical. Single-agent code ports easily. Multi-agent teams must be rethought from event-driven chat to data-flow graphs (migration guide).
  • MAF scope. It is single-process for now (no distributed runtime yet), many features are still preview, and it is Azure/Foundry-centric in samples and defaults (migration guide, MAF 1.0).
  • Responsibility stays with you. Microsoft states that responsible-AI mitigations and data-flow compliance with third-party systems are the builder’s job (MAF overview).

When to use it, and when not

  • AutoGen: only to maintain or migrate existing systems, or to read it as a reference design for conversation-based multi-agent patterns and Magentic-One.
  • Agent Framework: for .NET shops, Azure / Azure AI Foundry deployments, teams already on Semantic Kernel, or anywhere you need typed, checkpointed workflows with human-in-the-loop and a range of orchestration patterns behind one API.
  • Neither: if a function or single call suffices. Microsoft’s own rule is to write a function instead of an agent when you can (MAF overview). For Python-first graph control, LangGraph and Google ADK 2.0 are peers. For role/task crews, use CrewAI. For thin agent loops, use the OpenAI Agents SDK or the Claude Agent SDK.

Sources

Suggest changes →