JiuwenSwarm – Huawei's Open-Source AI Agent Unified Work Platform

1. What is JiuwenSwarm

JiuwenSwarm is an open-source AI Agent platform jointly developed by Huawei's 2012 Lab, Huawei Cloud, and the Xiaoyi terminal team. It is also the first open-source unified AI workbench designed for HarmonyOS PCs. The platform employs a multi-agent collaborative architecture, supporting one-stop task execution across various scenarios such as office work, programming, and entertainment. It innovatively introduces the HITS human-agent collaboration paradigm, enabling humans and AI agents to work together as a team to accomplish complex tasks. The platform is fully open-source and supports deployment on multiple platforms, including Windows, Mac, HarmonyOS, and Ubuntu.

Image source: Official article
Image source: official article

Technical positioning and domain: JiuwenSwarm belongs to the field of multi-agent collaboration (Multi-Agent Collaboration), and is positioned as a cross-scenario unified AI workbench. Unlike traditional single-agent tools, it combines multiple specialized agents into a team through a cluster model to collaboratively complete complex tasks, offering a unified entry point for scenarios such as office automation, intelligent programming, and interactive entertainment. Its unique feature is native compatibility with the HarmonyOS PC ecosystem, filling the gap in the domain of unified AI workbenches on this platform.

Development background: This platform was jointly developed by Huawei's 2012 Lab, Huawei Cloud, and the Xiaoyi terminal team. The 2012 Lab is Huawei's leading-edge technology research institution, with long-term investment in AI foundational capabilities; Huawei Cloud provides cloud computing and AI service support; and the Xiaoyi terminal team has accumulated rich experience in on-device AI interaction. The collaboration aims to integrate Huawei's technical advantages in AI, cloud computing, and terminal ecosystems to build an AI Agent infrastructure tailored for the HarmonyOS ecosystem.

Core value: JiuwenSwarm addresses the current fragmentation of AI tools by unifying AI capabilities across office, programming, and entertainment scenarios into a single platform. Its proposed HITS (Human-In-The-Swarm) human-agent collaboration paradigm transforms humans from task supervisors into team collaborators, enabling joint reasoning, decision-making, and execution with AI agents. This significantly improves the quality and efficiency of complex task completion. Additionally, its full open-source strategy reduces the barriers for enterprises and individuals to build their own agent platforms.

2. Key Features

• Cluster Mode for Office Work: Users can form a multi-agent team with just one sentence, and the system automatically assigns roles (such as researcher, architect, content creator, integrator) to proceed with tasks in parallel. According to official information, this mode can generate a 200-page PPT within 20 minutes, significantly reducing the time required to prepare presentation materials, especially suitable for urgent content creation scenarios.

• Intelligent Programming Development: In Code mode, agents can autonomously plan ArkTS code development tasks, supporting parallel front-end and back-end development, branch merging, and requirement additions. Developers only need to describe the requirements, and the agent can complete the entire process from coding to compilation and execution. This is particularly suitable for rapid iteration of HarmonyOS native applications, lowering the development barrier.

• Human-Agent Collaborative Gaming: The platform includes interactive games such as Gomoku and Werewolf, with agents capable of serving as hosts or AI teammates. The HITS paradigm is demonstrated here through joint reasoning and gaming between humans and agents, enhancing the immersion and social aspects of the game. It also provides an experimental environment for human-agent collaboration research, showcasing the potential of multi-agent systems in non-task scenarios.

• Cross-Device Integration: The platform supports deep integration with applications such as Feishu and Xiaoyi. Users can remotely activate the agent team on HarmonyOS PCs via their mobile phones, enabling cross-device task scheduling. For example, triggering a PPT generation task through a Feishu message while away from the office, with the results automatically sent back, forming a seamless cross-device workflow.

• Multi-Platform Deployment: Native installation packages are provided for operating systems such as Windows, Mac, HarmonyOS, and Ubuntu. Users can choose the appropriate version based on their device. The HarmonyOS PC version has been deeply optimized for the HarmonyOS system, ensuring consistency with the system's UI and interaction logic, delivering a plug-and-play experience.

• Agent Orchestration and Customization: The platform provides a visual agent orchestration interface, allowing users to customize the roles, tools, and collaboration processes of agents. Enterprise users can create dedicated agent teams based on their business needs to automate business processes. This feature reduces the development barrier of multi-agent systems, enabling non-technical users to configure complex collaboration logic.

3. How to Use

1. Environment Requirements and Prerequisites: JiuwenSwarm supports Windows 10/11, macOS 12+, HarmonyOS 3+, and Ubuntu 20.04+ operating systems. In terms of hardware, it is recommended to have at least 8GB of memory and a dedicated graphics card (for local model inference). However, the platform also supports connecting to cloud AI services to reduce local computational demands. Before use, ensure that the corresponding runtime environment for the platform is installed (e.g., HarmonyOS devices require HarmonyOS system support), and prepare a valid AI model API key.

2. Download and Installation: Visit the JiuwenSwarm official website https://openjiuwen.com/jiuwenswarm, go to the download page, and select the installation package for your operating system. After downloading, complete the installation following the system instructions. For example, on Windows, double-click the installation package and follow the wizard to complete the installation. During the first launch, the system will perform initial configuration, including agreeing to the license terms and selecting a data storage path.

3. Configure AI Model Services: After installation, open the JiuwenSwarm client and go to the settings interface. Under the "Model Services" option, users need to connect to an AI model to provide the agent with reasoning capabilities. Supported AI models include Huawei Cloud ModelArts, OpenAI API, and locally deployed open-source models (such as LLaMA, Qwen, etc.). It is recommended to choose a suitable model based on the complexity of the task. For office scenarios, a medium-sized model can balance speed and quality, while for programming scenarios, a model with strong code capabilities is recommended.

4. Select Work Mode and Initiate Tasks: Based on the task type, switch to "Office," "Code," or "Entertainment" mode on the main interface. The system will automatically match the pre-set agent team templates. Users can input a task description (e.g., "Generate a 20-page PPT on market analysis"), and the agent team will automatically break down the task and collaborate to advance it. During task execution, users can intervene at any time through the HITS interface to review intermediate results, adjust the direction, or add additional requirements, achieving true human-machine collaboration.

5. Monitoring and Collaboration: The platform provides a real-time task dashboard to display the progress and output of each Agent. Users can interact with specific Agents through the chat interface to provide additional instructions or feedback. After task completion, the results will be automatically saved to the local device or a designated cloud storage location. For long-running tasks, the platform supports background execution and completion notifications, so users do not need to keep the application in the foreground continuously.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: If you are a HarmonyOS PC user or wish to handle office, programming, and entertainment tasks on a unified platform, JiuwenSwarm is the best choice. Its out-of-the-box installation package and HITS human-machine collaboration paradigm are particularly suitable for non-technical users to get started quickly, as well as for complex scenarios requiring close collaboration between humans and machines. For enterprise users, the full open-source strategy supports private deployment, enabling the construction of internal agent platforms.

For developers focusing on programming and general agent orchestration, AutoGen and MetaGPT offer more flexible configuration capabilities and mature community ecosystems. AutoGen has deep expertise in agent-to-agent communication and tool usage, making it suitable for building complex multi-agent systems. MetaGPT excels in automating software development processes by simulating roles in a software company, especially suitable for teams requiring structured role division. CrewAI is known for its simple API and rich plugin system, making it ideal for scenarios requiring rapid workflow customization. However, its visual capabilities are weaker, making it more suitable for users with some programming experience.

6. Editor's Summary

As a significant open-source project by Huawei in the AI Agent domain, JiuwenSwarm's technological innovation is primarily reflected in three aspects: first, it is the first to deeply integrate a multi-agent collaborative architecture with the HarmonyOS PC ecosystem, filling the gap in the platform's AI unified workspace capabilities and providing foundational infrastructure for the practical implementation of AI capabilities within the HarmonyOS ecosystem; second, it introduces the HITS human-machine collaboration paradigm, transforming humans from task supervisors into team collaborators. This design is unique among current Agent frameworks, offering a more natural interaction model for human supervision and intervention in complex tasks; third, through cluster mode parallel acceleration, it has achieved significant improvements in office and development efficiency, such as the ability to generate a 200-page PPT in 20 minutes, which is quite remarkable among similar tools and demonstrates the practical value of the multi-Agent parallel architecture.

In terms of practical value, JiuwenSwarm reduces the usage threshold for multi-agent systems, allowing users to assemble Agent teams and complete complex tasks without writing any code. Its cross-scenario coverage (office, programming, entertainment) makes it a true "unified workspace," ideal for users who frequently switch between tasks, avoiding the fragmented experience of switching between different tools. For enterprise users, the full open-source strategy means they can deploy it privately and perform secondary development based on it, meeting internal business process automation needs while avoiding dependency on a single vendor.

In terms of target users, JiuwenSwarm mainly caters to HarmonyOS PC users, office efficiency seekers, HarmonyOS application developers, and AI Agent researchers. For non-technical users, its out-of-the-box installation and intuitive interface lower the usage barrier. For developers, the open-source code and orchestration interfaces provide room for expansion, enabling the construction of more complex agent systems based on it. However, currently, it mainly supports the ArkTS language in programming scenarios, and the richness of office scenario templates still needs improvement. These areas will be key directions for future development.

In terms of future development potential, as the HarmonyOS PC ecosystem matures, JiuwenSwarm has the potential to become a core entry point for AI capabilities within that ecosystem. If Huawei continues to improve documentation, expand the community, and introduce more scenario templates and language support, the platform could establish a unique competitive advantage in the AI Agent field. Additionally, the promotion of the HITS paradigm may influence the design philosophies of other Agent frameworks, driving human-machine collaboration from a "command-execution" model toward a "collaboration-collaborative creation" model.

7. Application Scenarios

• Emergency Report Generation: While on the move, users can send a command via their mobile Feishu app to remotely activate the JiuwenSwarm agent team on a HarmonyOS PC. The agents automatically gather information, organize the structure, and create a PowerPoint presentation, generating a 200-page report within 20 minutes that can be used directly for presentations. This scenario fully leverages cross-device collaboration and parallel multi-agent processing to address the inefficiency of mobile devices in handling complex documents.

• Rapid Development of HarmonyOS Native Applications: Developers can simply describe application requirements in natural language, and JiuwenSwarm's Code mode automatically plans tasks and invokes the ArkTS code generation, compilation, and installation process. Front-end and back-end agents work in parallel, supporting additional requirements and branch merging, significantly shortening the cycle from requirement to application deployment. This scenario is especially suitable for prototyping and rapid iteration of HarmonyOS applications.

• Multiplayer Online Collaboration Games: When organizing games like Gomoku or Werewolf, JiuwenSwarm agents can act as hosts (responsible for rule explanations and process control) or AI teammates (collaborating with human players in reasoning and voting). The HITS paradigm enhances the interactivity and unpredictability of the game process, making it ideal for team-building or social entertainment, while also providing rich experimental data for human-AI collaboration research.

• Cross-Device Intelligent Office Work: Users can initiate tasks via the Xiaoyi voice assistant on their mobile phone, and the task is automatically transferred to JiuwenSwarm on the HarmonyOS PC for execution. For example, a voice command such as "Organize last week's meeting minutes and generate a to-do list" will be automatically completed by the agent team on the PC, with the results synchronized back to the phone. This scenario enables seamless collaboration between mobile and PC devices, enhancing the efficiency of mobile office work.

• Enterprise-Level Multi-Agent Scheduling: Enterprises can build private deployments based on the open-source code of JiuwenSwarm, customizing agent roles and collaboration processes. For example, in a customer service scenario, agents can be set up for information collection, issue classification, and solution provision, working together to handle user requests and automate business processes. Enterprises can also use a visual orchestration interface to quickly adjust the structure of the agent team according to changing business needs.

8. FAQ

Q: Does JiuwenSwarm require an internet connection to function?
A: The JiuwenSwarm client itself is a local application, but the reasoning capabilities of the agents depend on the configured AI model services. If using a cloud API (such as Huawei Cloud ModelArts or OpenAI), an internet connection is required; if using an open-source model deployed locally (such as through Ollama or vLLM), it can operate in an offline environment. It is recommended to choose an appropriate model integration method based on privacy requirements and network conditions.

Q: How can I configure AI models for JiuwenSwarm?
A: In the client's "Settings" -> "Model Services" section, users can add API endpoints. Supported services include OpenAI-compatible interfaces, Huawei Cloud ModelArts, and local model services. When configuring, you need to provide the API address, key, and model name. It is recommended to select a model appropriate for the task complexity; for office scenarios, GPT-4 or models with equivalent capabilities are recommended, while for programming scenarios, code-optimized models (such as CodeLlama, DeepSeek-Coder) are suggested.

Q: Does JiuwenSwarm support private deployment?
A: Yes. The JiuwenSwarm code is fully open-sourced on GitHub, and enterprises can download the source code for private deployment, modifying agent logic and adding custom tools as needed. When deploying, you must configure the model services and runtime environment yourself. It is recommended to use Docker for containerized deployment to simplify dependency management. For specific deployment instructions, refer to the README and Wiki in the project repository.

Q: What are the system requirements for the HarmonyOS PC version of JiuwenSwarm?
A: The HarmonyOS PC version requires HarmonyOS 3.0 or higher. It is recommended that the device has at least 8GB of memory and 10GB of storage space. This version has been optimized for UI and interaction specifically for the HarmonyOS system, supporting multi-window and task switching features. Note that the HarmonyOS PC version is currently in an early stage, and some features may change with future system updates.

Q: How does JiuwenSwarm differ from frameworks like AutoGen and MetaGPT?
A: JiuwenSwarm is an AI unified workspace designed for end users, offering a ready-to-use client and a visual interface. In contrast, AutoGen and MetaGPT are more developer-oriented frameworks that require programming configuration. Additionally, JiuwenSwarm's HITS human-agent collaboration paradigm is a unique advantage, and it natively supports the HarmonyOS PC ecosystem. In terms of scenario coverage, JiuwenSwarm integrates office, programming, and entertainment functions into one platform, whereas other frameworks typically focus on programming or general orchestration.

Q: Can I use all of JiuwenSwarm's features on non-HarmonyOS devices?
A: JiuwenSwarm supports Windows, Mac, and Ubuntu, and on these platforms, the office, programming, and entertainment modes can be used normally. However, features deeply integrated with the HarmonyOS ecosystem, such as cross-device coordination (e.g., remote wake-up of the PC from a phone, deep integration with Feishu or Xiaoyi), may not be fully available on non-HarmonyOS devices. When using JiuwenSwarm on non-HarmonyOS devices, it is recommended to focus on localized features, and cross-device coordination can be awaited for future support on other platforms.

9. Project Links

• Project Website: https://openjiuwen.com/jiuwenswarm
• GitHub Repository: https://github.com/openJiuwen-ai/jiuwenswarm/