Full-Code · Orchestration Frameworksactive

Qwen-Agent

Type: full-code · Vendor: Alibaba Qwen team (QwenLM) · Language: Python · License: Apache-2.0 · Status: active · Status in practice: emerging · First released: 2023-09-22

Links: homepage docs repo

Provide a Python framework for building LLM applications that exercise the instruction-following, tool-use, planning, and memory capabilities of Alibaba's Qwen models, with built-in support for function calling, MCP, a Docker-sandboxed code interpreter, and RAG over very long documents.

Description. Qwen-Agent is the Apache-2.0 framework maintained by Alibaba's Qwen team for building agentic applications on top of Qwen models (Qwen >= 3.0). It ships in two languages — the English README and a parallel 中文 README_CN — and centres on three first-class capabilities advertised in the GitHub repo description: Function Calling, MCP, Code Interpreter, RAG. Distinctive characteristics: tight coupling to Qwen's native tool-call template (parallel function calls supported by default), Mandarin-first documentation with full Chinese-language abstractions and prompts, and a bundled fast-RAG pipeline aimed at super-long-document QA rather than generic vector-store glue.

Agent loop shape. Class-hierarchy-based imperative loop. The Agent base class defines an iterator interface; subclasses compose function calling, ReAct-style reasoning, and RAG: FnCallAgent is the function-calling base, ReActChat wraps the loop in a ReAct Thought/Action/Observation prompt, Assistant layers RAG over function-calling. Multi-agent shapes (group_chat, group_chat_auto_router, router) live in sibling modules.

Primary use cases

Qwen-model-native function calling and tool use
RAG over super-long documents
ReAct-style agents with sandboxed code execution
agentic apps with Mandarin-first prompts and docs

flowchart TD USER[User query] --> AGENT_BASE[class Agent base — _run iterator] AGENT_BASE -.->|inherits| FN[class FnCallAgent function calling] FN -.->|inherits| REACT[class ReActChat Thought/Action/Observation prompt] FN -.->|inherits| ASSIST[class Assistant RAG + function calling] REACT -->|tool call| TOOLS[(Tools built-in or @tool)] ASSIST -->|tool call| TOOLS TOOLS --> CI[Code Interpreter Docker-sandboxed] TOOLS --> MCP[MCP servers] ASSIST --> RAG[(Long-document RAG doc_qa subpackage)] RAG --> KB[(Knowledge base files)] AGENT_BASE -.->|backed by| QWEN[Qwen >= 3.0 LLM]

Key concepts

class Agent (docs) — Base class: 'An agent can receive messages and provide response by LLM or Tools. Different agents have distinct workflows for processing messages and generating responses in the _run method.'
class FnCallAgent (docs) — Function-calling base agent integrating LLM and tool use.
class Assistant → agentic-rag (docs) — Default agent class integrating RAG and function calling.
class ReActChat → react (docs) — ReAct-format agent carrying the embedded Thought/Action/Action Input/Observation prompt template.
Code Interpreter tool → code-as-action (docs) — Docker-sandboxed code execution as a built-in tool the agent invokes autonomously.