While the industry is still arguing over whether Loop Engineering (polishing a single agent's execution loop) has peaked, and whether Graph Engineering (designing multi-agent organizations as programmable graphs) is old wine in a new bottle, the Huawei-backed open-source AI agent platform openJiuwen put another paradigm on the table back in April 2026: Coordination Engineering — how an agent team forms, how it negotiates dynamically at runtime, and how proven collaboration experience gets distilled into reusable assets that keep evolving. JiuwenSwarm is that paradigm delivered as open-source engineering (Apache-2.0). The latest update merged Work mode and Code mode into a unified workbench, and launched a new human-agent collaboration paradigm, HITS (Human in the Swarm) — the human no longer only directs from outside the team, but walks straight into the agent team to work alongside, even to play games together.

GitHub repository: https://github.com/openJiuwen-ai/jiuwenswarm AtomGit mirror: https://atomgit.com/openJiuwen/jiuwenswarm

Core Positioning

JiuwenSwarm's tagline is a down-to-earth one: "Understands Your Intent, Evolves Autonomously — Swarm Collaboration for Complex Tasks". What it solves is not "how well a single agent runs" but "how multiple agents coordinate + how humans enter that coordination." The latest major release (v0.2.0, 2026-05-18) merged the previously split "Work mode / Code mode" into a unified workbench: writing documents, doing development, and playing games no longer require switching back and forth — everything gets dispatched, advanced, and delivered in one place.

Eight Core Capabilities

Three execution modes switch as needed: Plan (step-by-step) / Performance (flexible parallelism) / Swarm (the Leader coordinates multi-agent collaboration; the default).

HITS: The Human No Longer Stands Outside the Team

The most imaginative part of this update is HITS (Human in the Swarm), set against the traditional HOTS (Human on the Swarm):

• HOTS: the human stands outside the team, directing and clicking approve — the default mode of all past human-agent collaboration
• HITS: the human walks into the swarm with an identity, joining the conversation as a Human Agent to reason, negotiate, and deliver alongside the AI

It is not "a human watching AI work"; both sides sit at the same table. The official examples already cover three scenario categories: office work, programming, and entertainment.

Use Cases (From Official Examples)

Office Scenarios

• Summon a report-material team on the train: send one sentence via Feishu on your phone; the AI runs on your computer while you review the draft, send it back if unsatisfied, pull in an expert agent or a polish agent mid-stream, and round-by-round acceptance gets it to usable.
• 200 pages of high-quality PPT in 20 minutes: pick swarm mode, enter the topic, audience, and style; multiple agents research / structure / draft / integrate in parallel — the formerly serial "research — outline — fill in pages" becomes a task chain that advances concurrently.

Programming Scenarios

• Building its own arcade Tank Battle: in Code mode the swarm forms a team — gameplay + UI, frontend, backend, testing, and one Leader each — develops on branches in parallel, then merges to main; after a few rounds of collaboration it actually runs. Mid-development you can @ a member to add a requirement on the fly.
• A fitness calorie-tracking app: describe the requirement in Code mode and the system plans tasks autonomously, completes the project, and produces a working app straight out of the build.

Life & Entertainment (HITS at Its Most Visible)

• Gomoku: an agent acts as host, invites two human players in, maintains the rules, reminds players to move, and can recommend positions — the agent doesn't just keep its head down working; it can step up as the organizer.
• Werewolf: you play villager / seer / werewolf, and the AI players read your statements, deduce identities, judge whether you're genuinely analyzing or muddying the waters, and vote in the key rounds. You can @ a member to whisper and trade secret signals — the relationship between human and AI becomes a bluffing game at the same table.

Installation and Getting Started

The JiuwenSwarm repo's tagline says "up and running in 3 commands," and it's true.

Option 1: pip Install (Recommended)

# Users in China: use the Tsinghua mirror
pip install jiuwenswarm -i https://pypi.tuna.tsinghua.edu.cn/simple

# First-time initialization
jiuwenswarm-init

# Start the service, then open http://localhost:5173 in a browser
jiuwenswarm-start

💡 Tip: jiuwenswarm-start launches a local service with the frontend on port 5173 by default; once model service configuration is done (Huawei Cloud MaaS / OpenAI / DeepSeek / DashScope / SiliconFlow / OpenRouter / local), you can start dispatching tasks.

Option 2: TUI Terminal Interface

pip install jiuwenswarm-tui -i https://pypi.tuna.tsinghua.edu.cn/simple

# Run in a new terminal
jiuwenswarm-tui

Option 3: One-Click Desktop Install

Download page on the official site (covering Windows 10/11, macOS Intel & Apple Silicon): https://www.openjiuwen.com/download

Option 4: From Source

Requires Python ≥ 3.11; supports Windows / macOS / Linux.

git clone https://github.com/openJiuwen-ai/jiuwenswarm
cd jiuwenswarm

Working Examples of the Three Execution Modes

Just feed it input text and it runs:

Swarm mode (default, multi-agent collaboration):

Conduct an in-depth research on the new energy vehicle industry and generate an analysis report.

Planning mode (step-by-step execution):

Help me build a personal blog website — from tech selection, page design, frontend and backend development to deployment, create a complete development plan and execute step by step.

Performance mode (flexible parallelism):

Check today's weather in Beijing, and recommend 3 books about artificial intelligence.

Roadmap (Worth Watching)

Two items planned for 2026-07 bear directly on turning HITS into a production reality:

• Swarmflow Stateful Operators: nodes support human intervention and state memory, with humans acting as stateful operators who approve / correct / take over.
• Team Mode Same-Session Feishu Integration: Feishu users join the same session as Human Agents — human-AI mixed squads.

If both land on schedule, "humans joining the conversation directly as agents inside Feishu" becomes HITS' first product form that's genuinely usable day to day.

Use Cases and Limits

A good fit for:

• Teams and individuals who need multiple agents collaborating on complex tasks (long reports, complete apps, research work)
• Researchers and developers interested in "human-agent collaboration" who want to experience the HITS paradigm firsthand
• Developers in China: Tsinghua mirror + Huawei Cloud MaaS + HarmonyOS support keep the deployment bar low

Current limitations:

• The Skill Hub ecosystem is still under construction (it only officially kicked off with the 2026-05-21 roundtable livestream)
• Cross-machine distributed deployment takes some engineering skill to tune
• The desktop client currently targets Win/Mac mainly; HarmonyOS support is in progress

Final Thoughts

JiuwenSwarm's landing point is clear: tasks are no longer handed to a single agent but to a squad, and humans no longer stand outside the squad — they team up with the agent group. Add the integrations with Feishu, Celia, and other surfaces, and people can keep advancing tasks across different devices and collaboration environments. If you're looking for an open-source, evolvable multi-agent platform designed with "humans joining the team" as a first-class citizen, JiuwenSwarm is worth running pip install jiuwenswarm on today and trying end to end.

Repository: https://github.com/openJiuwen-ai/jiuwenswarm Mirror: https://atomgit.com/openJiuwen/jiuwenswarm Download: https://www.openjiuwen.com/download