AgentScope vs AutoGen

AgentScope 2.x is an actively released Apache-2.0 framework: a ReAct agent with a permission system, sandboxed workspaces, experimental team pipelines and an agent service, with MCP, A2A and AG-UI documented. AutoGen offers group-chat teams on an actor-style core but is in maintenance mode with no new features. For new work, AgentScope is the maintained option.

Facts side by side

AgentScope vs AutoGen. Data as of 2026-09-30.
Fact AgentScope AutoGen
Type Framework Framework
Languages / SDKs Python, TypeScript, Java Python, .NET
License Apache-2.0 CC-BY-4.0
Pricing model Open source, free Open source, free
Orchestration pattern Supervisor Blackboard
GitHub stars 32,617 (as of 2026-09-30) 61,238 (as of 2026-09-30)
GitHub forks 3,592 9,271
Last push 2026-09-30 2026-04-15
Latest release v2.0.9 python-v0.7.5
Repository agentscope-ai/agentscope microsoft/autogen
Website agentscope.io microsoft.github.io
Documentation docs.agentscope.io microsoft.github.io
Last verified 2026-09-30 2026-09-30
MCP support Yes (checked 2026-09-30) Yes (checked 2026-09-30)
A2A support Yes (checked 2026-09-30) Unknown (checked 2026-09-30)
AG-UI support Yes (checked 2026-09-30) Unknown (checked 2026-09-30)
Install verified 2026-09-30 2026-09-30

Choose AgentScope if

  • You want a project with current releases (v2.0.9 on 2026-09-28).
  • You need per-tool permission rules with allow, deny or ask, and tools that run in Docker, E2B, Kubernetes or other workspaces.
  • You need A2A or AG-UI, or TypeScript and Java implementations.

Choose AutoGen if

  • You have existing AutoGen AgentChat teams (SelectorGroupChat, Swarm, MagenticOneGroupChat) and only need fixes.
  • You want peer-style group chat where agents share one conversation; AgentScope's TeamPipeline members never talk to each other.
  • You use AutoGen Studio for no-code prototyping.

Migration notes

AutoGen's AssistantAgent maps to AgentScope's Agent with a Toolkit. AutoGen group chats do not map directly: TeamPipeline is leader-and-members (members run in their own contexts and return final replies only), and GoalPipeline is an executor-verifier loop. Shared-conversation debates must be rebuilt by hand. Human input changes from a blocking UserProxyAgent to permission events (RequireUserConfirmEvent) that park and resume a reply. Caveats on the AgentScope side: the pipeline module is marked experimental, 2.x APIs differ from 1.x and should not be mixed, Python 3.11 or newer is required, and A2AAgent is a proxy without tools, permissions or human-in-the-loop for the remote agent. If you prefer to stay in Microsoft's ecosystem, the documented path from AutoGen is Microsoft Agent Framework.

FAQ

Is AgentScope a fork of AutoGen?

No. It is a separate project (agentscope-ai/agentscope) with its own architecture.

Which is actively developed?

AgentScope publishes frequent releases. AutoGen's README states maintenance mode with no new features.

Which supports more protocols?

AgentScope: MCP, A2A and AG-UI, all with evidence on its tool page (AG-UI through the agent service middleware). AutoGen: MCP client only; A2A and AG-UI are unknown.