AgentScope, an open-source framework developed by Alibaba’s Tongyi Lab, is drawing renewed attention from developers as the company expands it into what it calls agent-oriented programming: a way of building software in which a developer describes a goal and a team of AI agents is assembled to carry it out.
The framework comes from the same Alibaba research organization responsible for the Qwen family of models. It is available on GitHub and PyPI under an open-source license, and Alibaba has released a 1.0 version, documented in an August 2025 research paper, followed by a series of 2.x updates and a separate Java edition aimed at enterprise developers.
What is in the box
AgentScope’s pitch is that the pieces needed for a working agent system ship together rather than having to be assembled from separate libraries. According to the project’s documentation, it includes agent memory, retrieval-augmented generation (RAG) for grounding answers in documents, structured reasoning, and native support for the Model Context Protocol (MCP), the open standard for connecting agents to external tools and data sources.
It also includes AgentScope Studio, a visual interface for designing, running and tracing agents, so that developers can watch how agents reason, which tools they call and where they fail. Agents built with the framework can retain context across steps, revise their plans when a tool call fails and hand work to other agents in a multi-agent configuration.
Built for deployment, not only demos
Alibaba positions AgentScope as a production framework rather than a prototyping toolkit. The documentation describes running agents as services locally, in the cloud or in Kubernetes clusters, with support for asynchronous execution, real-time interruption and evaluation tooling. The framework is model-agnostic, working with Alibaba’s own Qwen models as well as APIs from OpenAI, Anthropic, Google and others.
That places it in competition with a crowded field of agent frameworks, including LangGraph, Microsoft’s AutoGen and Semantic Kernel, CrewAI and OpenAI’s Agents SDK. AgentScope’s distinguishing features are its emphasis on multi-agent orchestration, the integrated visual studio and its close ties to the Qwen ecosystem, which has a large developer base in China and growing adoption elsewhere.
Why it matters
Agent frameworks are becoming the layer where much of the practical work of applying AI happens, because most real tasks require a model to take several steps, use tools and recover from mistakes rather than answer a single question. Alibaba’s investment in AgentScope, alongside its models, signals that the company wants to own that layer as well, and to make it easy for developers who already use Qwen to build agents that can be deployed inside business systems.
