joyagent-jdgenie

JoyAgent-JDGenie is the industry's first open-source, high-completeness, lightweight general-purpose multi-agent product, open-sourced by the relevant team at JD.com. Its goal is to solve the "last mile" problem of quickly building multi-agent products.

[Purpose] Currently, related open-source agents are mainly SDKs or frameworks, and users still need to do further development based on them, so they cannot be used out of the box directly. JoyAgent-JDGenie is an end-to-end multi-agent product that can directly answer or solve the input query or task. For example, if the user queries "give me an analysis of the recent trend of the US dollar and gold," JoyAgent-Genie can directly provide a web-page or PPT report document.

[Main Capabilities]

• End-to-end complete multi-agent product, ready to use out of the box, and supports secondary development.
• Agent framework protocol, supporting multiple agent design patterns (react mode, plan and executor mode, etc.).
• Multi-agent context management.
• High-concurrency DAG execution engine, pursuing ultimate execution efficiency.
• Sub-agents and tools are pluggable: multiple sub-agents and tools are preset.
• Multiple file delivery formats: html, ppt, markdown.
• Plan and tool invocation, RL optimization iteration.
• Full-link streaming output.
• multi-level: work level and task level.
• multi-pattern: plan and executor mode and react mode.
• Iteratively generate new tools based on existing tools, rather than directly generating new tools from 0 to 1 (reducing the generation of erroneous tools).
• Implicitly decompose existing tools into atomic tools, and automatically combine them into new tools based on atomic tools and large models (without spending manpower to predefine and decompose atomic tools).
• The agent product JoyAgent-JDGenie has been open-sourced as a whole, including the frontend, backend, framework, engine, and core sub-agents (report generation agent, code agent, PPT agent, file agent, etc.).

[Generalization Verification] To verify the generalization of JoyAgent-JDGenie, on the GAIA leaderboard it achieved an accuracy of 75.15% on the Validation set and 65.12% on the Test set, surpassing OWL (CAMEL), Smolagent (Huggingface), LRC-Huawei (Huawei), xManus (OpenManus), AutoAgent (The University of Hong Kong), and other well-known industry products.

[Lightweight Characteristics] JoyAgent-JDGenie is relatively lightweight. Unlike Alibaba's SpringAI-Alibaba, which needs to rely on related functions of the Alibaba Cloud Bailian platform (invoking LLMs based on the Bailian platform), and Coze, which relies on the Volcano Engine platform.

[Related Components]

• Multimodal Knowledge Management Platform: a RAG solution for multimodal unstructured knowledge, integrating parsing, retrieval, and generation capabilities, able to efficiently process complex documents and provide one-stop support for intelligent question answering and content generation. A comprehensive RAG platform for multimodal documents, providing a one-stop solution for intelligent question answering and content generation for complex documents (note: use the mrag branch).
• JoyDataAgent: Enterprise internal knowledge mainly includes structured tabular knowledge and unstructured knowledge. Therefore, for structured tabular knowledge, an out-of-the-box DataAgent capability has been built, mainly including the data governance DGP protocol, intelligent data querying, and intelligent diagnostic analysis. JoyDataAgent is the first open-source agent that includes the data governance DGP protocol, intelligent data querying, and intelligent diagnostic analysis (note: use the data_agent branch).

[Usage] Method 1: one-click service startup with docker. If you encounter problems during deployment, you can refer to the video: [Start the open-source agent application joyagent-genie with deepseek in 5 minutes - bilibili] https://b23.tv/8VQDBOK Method 2: Manually initialize the environment and start the services. Python environment setup: pip install uv cd genie-tool uv sync source .venv/bin/activate For a detailed manual guide, refer to Step by Step. Start all services directly via shell. For deployment, you can refer to the video: [joyagent-jdgenie deployment demo] https://www.bilibili.com/video/BV1Py8Yz4ELK/?vd_source=a5601a346d433a490c55293e76180c9d

[How to add your own MCP tool to JoyAgent-JDGenie]

• Add the mcp_server service in genie-backend/src/main/resources/application.yml; separate multiple servers with commas.
• In ui/.env, you can modify the path for frontend requests to the backend.
• For example, after adding the 12306 tool, plan a 7-day, 2-person trip from Beijing to Xinjiang in July and query related train ticket information. Genie will design the travel plan, then call the MCP tool to query ticket information, and finally output a report.

[Add a custom sub-Agent to JoyAgent-JDGenie]

• Implement the BaseTool interface, declaring the tool's name, description, parameters, and invocation method.
• In com.jd.genie.controller.GenieController#buildToolCollection, add the following code to introduce the custom Agent.

[Applicable scenarios] Suitable for scenarios that require quickly building and deploying general-purpose multi-agent products, such as automatically generating analysis reports (web or PPT versions), travel plan design and train ticket queries, intelligent Q&A and content generation for structured and unstructured knowledge within enterprises, etc. For new scenario features that users need to customize, simply mount the relevant sub-agents or tools to JoyAgent-Genie.

[Limitations] The materials do not explicitly list usage limitations. The project welcomes all good ideas and suggestions. If you want to become a contributor to the project, you can submit a Pull Request at any time. Before that, you need to read and sign the contributor agreement and send it to the email org.developer3@jd.com. Please read the Chinese version of the contribution guide and the English version of the contribution guide. For academic citation, please use the BibTeX provided by the project.