Qwen-Agent is a framework for building large language model applications that leverages the instruction-following, tool usage, planning, and memory capabilities of Qwen models version 3.0 and above. The framework serves as the backend infrastructure for Qwen Chat and provides developers with both atomic and high-level components for constructing AI agents capable of complex task automation.
The repository is classified across multiple domains including multi-agent systems, LLM agents, task automation, tool integration, AI planning, and orchestration. It addresses the need for frameworks that can coordinate autonomous AI systems handling complex tasks through structured agent architectures and tool interactions.
Core components of Qwen-Agent include LLM classes inheriting from BaseChatModel with built-in function calling capabilities, Tool classes inheriting from BaseTool for extensibility, and Agent classes for higher-level orchestration. The framework supports parallel function calls as its default tool-calling template and provides both built-in agent implementations like Assistant and the ability to develop custom agents by inheriting from the Agent base class.
The framework includes several specialized applications and features. A Code Interpreter tool operates within Docker containers to allow agents to autonomously write and execute code in isolated sandbox environments. Model Context Protocol (MCP) support enables integration with external tools from the open-source MCP server ecosystem. RAG capabilities are provided for question-answering over documents up to 1 million tokens, with solutions that reportedly outperform native long-context models on challenging benchmarks. BrowserQwen serves as a browser assistant built on the framework. The framework also supports reasoning content fields and includes a Gradio-based GUI interface for rapid deployment of agent demonstrations.
Recent development activity shows the project maintains active engagement with 539 tracked issues and pull requests. The median response latency for issues and PRs is 5.9 hours, though the mean extends to 325.8 hours, indicating variable response times across different items. Work in Progress items represent the most active label category with 13 tracked instances, followed by Bug reports with 6 instances and Documentation with 3 instances. Primary contributors tuhahaha and JianxinMa have driven the majority of activity with 376 and 332 events respectively, while maxgameone contributed 30 events.
The repository maintains overlapping contributors with vllm-project/vllm, langgenius/dify, and infiniflow/ragflow, indicating integration points and shared development communities across these related projects. Installation is available through PyPI as the stable version or directly from source for development versions. The framework supports both Alibaba Cloud's DashScope model service and self-hosted deployments using open-source Qwen models, with specific guidance for vLLM and Ollama deployment options.
Recent updates include support for Qwen3.5 with agent demos, the DeepPlanning evaluation benchmark, Qwen3-VL tool-calling with image capabilities, native API tool call interfaces, reasoning content field support, QwQ-32B tool-calling with parallel and multi-turn support, and Qwen2.5-Math demonstrations of Tool-Integrated Reasoning. The framework is distributed under the Apache License 2.0 and includes comprehensive documentation and example implementations for developers building agent-based applications.
