AgentScope 2.0 – Tongyi Lab's Open Multi-Agent Development Framework

1. What Is AgentScope 2.0

AgentScope 2.0 is Alibaba Tongyi Lab's open multi-agent framework focused on moving agents from experimental demos to stable production. It follows Agent-Oriented Programming—agent autonomy and organization over simple chat orchestration. Six production systems differentiate it from prototype-only frameworks: model fault tolerance, systematic permissions, redesigned context management, Middleware hooks, Workspace execution decoupling, and built-in Agent Service deployment.

Technical positioning and domain: Enterprise-grade MAS framework vs AutoGen/CrewAI—emphasis on long-run stability, safety, and service deployment—not just collaboration graphs.

Development background: Tongyi/Qwen team shipping agents in production saw model failures, context overflow, permission gaps, and deploy friction—AgentScope 2.0 targets those last-mile engineering issues.

Core value: Closes the gap between "runs once" and "runs reliably"—retries/fallback models, auditable event streams, tool/file/shell guardrails, structured memory compression, and one-step service exposure.

Technical characteristics: Stability (retry + fallback), security (permissions), transparency (event stream), extensibility (Middleware + Workspace), plus multimodal and MCP/A2A support.

2. Key Features

• Model fault tolerance: Unified Qwen, Anthropic, DeepSeek, Gemini, OpenAI, Grok, Moonshot, etc., with auto-retry and fallback model switching so long workflows survive single-provider outages.

• Transparent event stream: Content blocks for text, files, images, tool results, and model thoughts—stream start/delta/tool/user-confirm events for live debug, audit, and human-in-the-loop.

• Permission system: Rules over tool calls, file read/write, and shell exec—static rules + risk scoring → allow/deny/ask-user; blocks dangerous paths (/etc, /sys), sensitive files, and rm -rf-class commands.

• Context management: Structured compression keeps goals, state, findings, and next steps; truncates huge tool logs; read-cache before write to cut redundant IO—supports tens to hundreds of turns.

• Middleware: Inject logging, policy, safety, or dynamic context at model/plan/tool stages without forking core code—web-style middleware for agents.

• Workspace abstraction: Local, Docker, or E2B sandboxes behind WorkspaceBase—same agent code, different runtime; warm pools for parallel env spin-up.

• Agent Service: Built-in HTTP service with streaming and session resume—scripts become production APIs without bespoke FastAPI glue.

3. How to Use

1. Install (Python 3.9+, recommend 3.10/3.11):

   pip install agentscope
   

   Install docker or e2b SDK if using those workspaces.

2. Model config (config.json example):

   {
     "models": {
       "primary": {
         "model_type": "qwen",
         "config": {
           "model": "qwen-max",
           "api_key": "your-key",
           "retry_times": 3
         }
       },
       "fallback": {
         "model_type": "openai",
         "config": {
           "model": "gpt-4o-mini",
           "api_key": "your-key"
         }
       }
     }
   }
   

3. Define agent + tools:

   from agentscope.agents import ReActAgent
   from agentscope.tools import Toolkit
   
   def search_web(query: str) -> str:
       return search_result
   
   agent = ReActAgent(
       name="ResearchAgent",
       model_config_name="primary",
       toolkit=Toolkit([search_web])
   )
   

4. Permissions example:

   {
     "permissions": {
       "file_read": {
         "allow_paths": ["/tmp/**"],
         "deny_paths": ["/etc/**", "/sys/**"]
       },
       "command_exec": {
         "allowed_commands": ["ls", "cat", "grep"],
         "risk_level": "user_confirm"
       }
     }
   }
   

5. Workspace: Pick local (default), Docker image, or E2B template in config.

6. Run or serve:

   from agentscope.service import AgentService
   service = AgentService(agent=agent)
   service.start(host="0.0.0.0", port=8000)
   

   Or locally: agent.reply("Search latest AI news").

Notes: Dev local → test Docker → prod E2B for isolation; define permissions early; enable compression on long jobs.

4. Pros and Cons

5. Comparison with Similar Tools

Selection guidance: Fast demos → CrewAI/AutoGen. Hundreds of turns, sensitive FS/shell, standard API, multi-env → AgentScope 2.0. Deep .NET/Azure → Semantic Kernel.

6. Editor's Take

AgentScope 2.0's bet is systemic production engineering—not one killer feature. Fault tolerance + permissions + event streams + Middleware + Workspace + Service form a demo→prod loop competitors lack piecemeal.

Permissions and observability are enterprise milestones; teams stuck at "works in notebook" should evaluate migration cost vs rebuilding guardrails themselves.

Best for backend automation engineers, agent researchers productizing, and regulated industries (finance/health/gov). Beginners: start with CrewAI/AutoGen concepts first.

Backed by Tongyi/Alibaba with MCP/A2A tailwinds; open source should accelerate plugins and docs.

— innovation & production 5/5; ease & community ~3.5/5.

7. Use Cases

• Enterprise multi-agent workflows: Research → analyze → approve pipelines with permissions and audit trails.
• Voice agents: STT/TTS/real-time voice with interruptible event streams.
• Risky ops: DevOps agents with command/file guardrails and human confirm on medium/high risk.
• Long-horizon research: Competitive analysis with compressed memory over many tool rounds.
• Hybrid deploy: Same agent local → Docker CI → E2B prod for compliance.

8. FAQ

Q: vs AutoGen?
A: AutoGen optimizes conversational multi-agent demos; AgentScope optimizes stable, secure production deployment.

Q: Private deploy?
A: Yes—fully OSS; use Ollama/vLLM backends; local Workspace without cloud sandboxes.

Q: Prevent dangerous actions?
A: Configure path/command/tool policies; auto-block high risk, confirm medium risk.

Q: Multimodal support?
A: Image gen, TTS, voice in/out via unified content blocks; MCP/A2A for external media services.

Q: Context overflow?
A: Structured compression + tool log truncation + read-cache—designed for 100+ turn jobs.

Q: Production readiness?
A: Agent Service + Workspace + fault tolerance used internally at Tongyi—intended for real workloads.

9. Project Links

• Docs: https://docs.agentscope.io/zh/v2
• GitHub: https://github.com/modelscope/agentscope