Key points
- From framework to platform: AgentScope (open-sourced by Alibaba Tongyi Lab in Feb 2024, ~15k GitHub stars) was rebuilt from the ground up in v2 (May 2025) around an API-First architecture, aiming to be a multi-agent *operating system* with a permission layer, workspace, system calls, and session management — unlike AutoGen (conversation orchestration), LangGraph (state-machine workflows), or CrewAI (role-based task allocation), which are libraries, not platforms.
- MD-LAD meta-architecture: v2 describes agents across three dimensions — Platform (system config, keys, permissions via
agentscope.init()), App (agent structure, tools, workflows), and User (session instances, context, events). This "onion-style" layering enforces runtime isolation and makes multi-tenancy native rather than patched on. - Several documentation paths (workflow, pipeline, tutorials, use-cases) returned 404s — v2 is still iterating rapidly.
- The complexity is a double-edged sword: eight modules, layered permissions, and middleware hooks raise the learning curve; v2 targets engineering teams shipping agents to production, not quick prototypers.
- Agent Service's multi-tenant stability is unproven at scale; community feedback over the next six months will shape its trajectory.
- v1 users: migration, not an upgrade — but worthwhile for production needs.
- Framework shoppers: choose LangGraph for deterministic state-machine workflows, AutoGen for multi-agent negotiation, AgentScope v2 for long-running, permission-controlled, deployable applications (ops automation, customer service, research assistants).
- Alibaba ecosystem users: native integration with Qwen/DashScope and Alibaba Cloud infrastructure is a home-field advantage competitors won't replicate.
- Official docs: https://docs.agentscope.io/v2
- Changelog: https://docs.agentscope.io/v2/change-log
- GitHub: https://github.com/agentscope-ai/agentscope
- Tongyi Lab: https://tongyi.aliyun.com
- MD-LAD design: https://docs.agentscope.io/v2/building-blocks/agent
The eight building blocks
1. Agent: async ReAct with token-level streaming (TextEvents accumulate into a complete Msg), plus event-driven human-in-the-loop pause/resume integrated with asyncio.
2. Message & Event: dual-mode data system — Msg is the durable communication/persistence unit; Event is the ephemeral streaming unit for UIs. One reply() yields many events but exactly one persisted Msg.
3. Workspace: interchangeable execution environments — Local, Docker, or E2B (Firecracker microVM sandboxes) — switchable via config, integrated with permissions.
4. Permission System: rule engine combining role-based and resource-based control with allow, deny, and confirm actions; high-risk operations (shell execution, file writes) are blocked *before* execution pending user confirmation.
5. Model: unified ChatModelBase across OpenAI, Anthropic, Google, DashScope, Ollama; native JSON Schema structured output, multimodal messages, automatic retry and fallback-model switching.
6. Context: three-layer protection for long-running agents — compression (summarization), truncation (tool output limits), and reset — enabling runs of hours or days without hitting context limits.
7. Tool: ToolBase/FunctionTool plus native MCP (Model Context Protocol) integration; a Skill mechanism bundles tools with prompt templates for reuse (a skill marketplace is hinted at).
8. Middleware: five lifecycle hooks (pre_reply, pre_model, post_model, post_reply, on_error) enabling logging, filtering, and alerting without forking the source.
Agent Service: script to production
The flagship v2 addition is a FastAPI-based HTTP service layer providing per-user session persistence with reconnect support, SSE streaming, scheduled tasks, background task queues, and a clear resource hierarchy (User → Credential/Agent/Schedule → Session/Workspace). Developers can prototype locally and deploy the same code as a multi-tenant service.
What changed from v1
Architecture philosophy (framework → OS-like platform), API-First contracts, full RBAC+ABAC permissions, Event-Msg dual-mode streaming, three Workspace options, production-grade Agent Service, three-layer context management, native MCP, and built-in multi-tenancy. The transparency and controllability philosophy from v1 remains — but now on stronger foundations.