AgentScope
Apache 2.0 agent framework from Alibaba Tongyi Lab's SysML team. Version 2.0 folded the separate Runtime project into the core, so sandboxing, agent serving and observability now ship as one...
Overview
Alibaba's Tongyi Lab SysML team shipped AgentScope 1.0 in September 2025 and followed with a 2.0 line whose defining change was structural. The separately maintained Runtime project, carrying tool sandboxing, Agent-as-a-Service APIs and full-stack observability, was pulled into the core package. That merge is the story here. One dependency instead of three removes the most common source of version mismatch that plagued the 1.x era, and it means a production deployment no longer needs two release cadences kept in step.
The agent core is a ReAct loop with structured output, interruption, resume and batched tool calls. Around it sits a toolkit layer that treats Python functions, MCP servers and reusable skills the same way, a multi-agent orchestrator with leader-worker teams and task planning, and context management that compacts history and offloads large tool results without manual intervention. Sandboxed execution can target local processes, Docker, Kubernetes, E2B, OpenSandbox or Daytona. Model support is deliberately broad: DashScope, DeepSeek, Moonshot, OpenAI, Anthropic, Gemini, xAI and Ollama all work without adapters.
Two practical notes for anyone deploying this in China. Python 3.11 is the hard minimum, which rules out several older enterprise base images that are still in wide use. And separate Java implementations exist at agentscope-java and agentscope-runtime-java, both Apache 2.0, which is more useful than it sounds in an environment where the production backend is almost always Spring. The August 2026 releases added a terminal console for testing and debugging agents plus channel integrations for Feishu and Discord, the former being the one that matters for a mainland internal tool. Version v2.0.7.post1 is dated 28 August 2026 against 29,900 stars, so the pace is not slowing.
Key Features
- ✓ ReAct agent core with structured output, interruption, resume and batched tool calls
- ✓ Toolkit layer that treats Python functions, MCP servers and reusable skills uniformly, with coding tools built in
- ✓ Sandboxed tool and code execution targeting local processes, Docker, Kubernetes, E2B, OpenSandbox or Daytona
- ✓ Multi-agent orchestration with leader-worker teams and task planning
- ✓ Context management with automatic compaction, tool-result offloading and middleware-based context injection
- ✓ Model providers covering DashScope, DeepSeek, Moonshot, OpenAI, Anthropic, Gemini, xAI and Ollama, plus channel integrations for Feishu and Discord added in August 2026
Where it holds
- • Plain Apache 2.0 from Alibaba Tongyi Lab's SysML team, with no additional conditions in the licence file
- • Runtime, sandboxing and observability merged into the core package at 2.0, replacing three coordinated dependencies with one
- • Release pace is genuinely fast, with v2.0.7.post1 dated 28 August 2026 against 29,900 stars
- • Separate Java implementations exist and are also Apache 2.0, which matters more than it sounds in Chinese enterprise where the production backend is almost always Spring rather than FastAPI
Where it breaks
- • Python 3.11 minimum rules out a number of older enterprise base images still common in mainland deployments
- • The 1.x to 2.0 refactor broke APIs, so most tutorials and blog posts written before 2026 are now wrong. Filter search results by date
- • Third-party integration ecosystem is smaller than LangGraph's, so expect to write your own connectors for anything unusual
- • English documentation covers the core API adequately, but the design rationale and the more advanced examples are noticeably richer in the Chinese material
My Take
Version 2.0 folded the separately maintained Runtime project back into the core package, and that consolidation is what makes it worth a second look now. Alibaba's Tongyi Lab SysML team ships it under plain Apache 2.0 at 29,900 stars, Python 3.11 or newer, with sandboxed tool execution that can target local processes, Docker, Kubernetes, E2B or Daytona. Model support spans DashScope, DeepSeek, Moonshot, OpenAI, Anthropic, Gemini and Ollama out of the box, so nothing forces you onto Qwen despite the parentage. Releases land almost daily, with v2.0.7.post1 dated 28 August 2026.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Free and Apache 2.0. Install from PyPI as agentscope with Python 3.11 or newer. Sandboxed tool execution runs locally or in Docker at no cost, and hosted sandbox targets such as E2B or Daytona are billed by those providers, not by AgentScope. Java implementations at agentscope-java and agentscope-runtime-java are also Apache 2.0. Model API spend goes to whichever provider you point it at, typically DashScope for Qwen inside China.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where AgentScope holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to AgentScope
LangGraph
open-sourceLow-level orchestration framework for stateful, long-running LLM agents, with checkpointing, human-in-the-loop pauses and a managed deployment tier through LangSmith.
OpenAI Agents SDK
open-sourceMinimal open-source Python and TypeScript framework for building agents with handoffs, guardrails, sessions and sandboxed code execution, usable with non-OpenAI models.
OpenManus
open-sourceMIT-licensed open agent framework from the MetaGPT team, prototyped in three hours as the open answer to Manus. 58,100 stars and still pushing commits in August 2026.
LlamaIndex
freemiumOpen-source Python and TypeScript framework for RAG and agentic workflows, paired with LlamaCloud for document parsing, extraction and managed indexes.