AgentScope, an open-source agent framework from Alibaba's Tongyi Lab, was first released in February 2024. Its v1 gained roughly 15,000 GitHub stars in a year—not dominant amid the AutoGen, LangGraph, and CrewAI wave—but it survived on one principle: transparency and controllability. Prompts, API calls, memory state, and workflows were all visible, interruptible, and modifiable.
The problem: visibility is not the same as production readiness. v1 lacked unified permission control, complete streaming pipelines, long-run context protection, and a deployment layer. Developers who built agent pipelines locally had to build their own gateways, session management, and model fallback logic.
v2: An architecture rewrite
Released in May 2025, AgentScope v2 is not a feature increment but a ground-up rewrite, declared on the first documentation page as an API-First architecture. The stated ambition: a multi-agent operating system—with kernel, filesystem, permission layer, system calls, and process management (sessions)—where agent applications can be installed, run, uninstalled, and upgraded like ordinary software. Unlike AutoGen (conversation orchestration), LangGraph (state-machine workflows), or CrewAI (role-based task allocation), all of which are libraries, AgentScope v2 aims to be a platform.
Key points
- MD-LAD (Multi-Dimensional-Layered Agent Description): a three-layer model that acts as the meta-rule of the architecture:
- Platform: system-level config (model keys, permission policies, global workspace) via
agentscope.init() - App: agent structure, tool composition, workflow topology via
Agent,Tool,Pipeline - User: session-level instances (conversation history, uploads, preferences) via
Session,Context,Event - Eight building blocks with explicit interface contracts: 1. Agent: async ReAct with streaming output (
- Agent Service: a FastAPI-based HTTP layer offering per-user sessions with persisted state and reconnection, SSE streaming, scheduled tasks, background queues, and a clean resource hierarchy (User → Credential / Agent / Schedule → Session / Workspace). Scripts can be deployed as services without hand-written session management.
- For v1 users: this is a migration, not a smooth upgrade. APIs, concepts, and architecture have all changed, but production needs (permissions, sessions, fault tolerance, deployment) justify it.
- For framework selectors: LangGraph suits step-defined business workflows; AutoGen suits multi-agent negotiation and role-play; AgentScope v2 suits long-running, human-collaborative, permission-governed, production-deployed applications (ops automation, customer service, research assistants).
- For the Alibaba ecosystem: native integration with Qwen, DashScope, and Alibaba Cloud is a home-field advantage competitors don't replicate.
- AgentScope v2 documentation: https://docs.agentscope.io/v2
- Changelog: https://docs.agentscope.io/v2/change-log
- GitHub: https://github.com/agentscope-ai/agentscope
- Alibaba Tongyi Lab: https://tongyi.aliyun.com
- MD-LAD design doc: https://docs.agentscope.io/v2/building-blocks/agent
This is runtime isolation, not just code organization—an "onion" structure where each layer talks only to adjacent layers, making multi-tenancy a native design rather than a patch.
reply() returns a Msg assembled progressively from TextEvents) and event-driven human-in-the-loop pause/resume integrated with Python asyncio.
2. Message & Event: a dual-modality data system—Msg (id, timestamp, content, role) is the persistent unit; Event (text, image, chunk, stop) is the transient streaming unit. One reply produces many events but exactly one persisted Msg.
3. Workspace: Local, Docker, or E2B (Firecracker microVM sandboxes), switchable via config; permission rules integrate with the active workspace.
4. Permission System: role-based plus resource-based rules with allow, deny, and confirm actions—dangerous operations like execute_shell_command can be blocked pre-emptively.
5. Model: unified ChatModelBase across OpenAI, Anthropic, Google, DashScope, Ollama, with native JSON Schema structured output, multimodal messages, and automatic retry/fallback.
6. Context: three-layer protection—compression (summarization), truncation (of tool outputs), and reset—enabling runs of hours or days without hitting context limits.
7. Tool: ToolBase, FunctionTool (Python functions auto-registered), MCPTool (Model Context Protocol integration), plus a Skill mechanism bundling tools with prompt templates for reuse.
8. Middleware: five lifecycle hooks (pre_reply, pre_model, post_model, post_reply, on_error) enabling logging, input mutation, output filtering, and alerting without source changes.
v1 vs v2
| Dimension | v1 | v2 | |---|---|---| | Architecture | dev framework | OS-like platform | | API design | example-driven | API-First contracts | | Permissions | none | RBAC + ABAC | | Streaming | basic | full Event-Msg duality | | Workspace | local mainly | Local / Docker / E2B | | Deployment | scripts | Agent Service | | Context | simple truncation | compress + truncate + reset | | MCP | none | native | | Multi-tenancy | unsupported | native |
The unchanged core is the transparency philosophy—v1's MsgHub, real-time control, and tracing survive in stronger form.
Practical implications
An honest assessment
The documentation is detailed but incomplete—multiple paths (workflow, pipeline, tutorial, use-cases) returned 404s, indicating ongoing iteration. The complexity is a double-edged sword: eight modules, three permission layers, and dual data modalities raise the learning curve, making v2 heavy for rapid prototyping. Its target users are engineering teams shipping agent applications, not solo prototypers. And Agent Service's production claims (multi-tenancy, persistence, scheduling) still await validation at scale—15,000 stars are not 15,000 production deployments. Community feedback over the next six months will shape v2's trajectory.
Conclusion
AgentScope v2 is a leap from framework to platform: an attempted operating system for the multi-agent era. Whether it succeeds depends on documentation and ecosystem completion, production validation under real concurrency, and clearer differentiation against LangGraph and AutoGen. Regardless, Alibaba's Tongyi Lab has presented an architecture proposal worth taking seriously—one with genuine design philosophy in a landscape of increasingly homogeneous agent frameworks.