This week, a GitHub trending project called JiuwenClaw caught the attention of developers who ran it themselves. The common thread in their reviews: "It actually works as a team." Multiple agents split roles like humans, communicated with each other, and delivered a finished product — all without a person in the loop.

AgentTeam: Multi-Agent Collaboration Goes Open Source

The openJiuwen community announced that the latest version of JiuwenClaw supports AgentTeam, a feature where multiple agents autonomously divide and collaborate on tasks. The development team calls this "Coordination Engineering," positioning it as the next step beyond Harness Engineering (controlling a single agent). In a live demo, the team collaboration mechanism showed remarkable stability. Team members operated with clear roles and autonomous coordination, requiring no human intervention in the entire workflow. Specifically, when asked to analyze OpenClaw technology across 10 core dimensions, each dimension had a dedicated agent generating 20 slides. The 10 sets of slides were merged under a unified theme, producing a 200-page technical PPT. The entire process took under 20 minutes. The project repository is available on GitHub.

What Used to Require a Human Handler

Previously, multi-agent collaboration was limited to simple task distribution and result aggregation. Dependency management and priority adjustments between agents required direct human intervention. JiuwenClaw's AgentTeam changes that structure. A Leader Agent orchestrates the entire team. The leader claims tasks, executes them, marks them complete, and unblocks subtasks. Team members communicate through two parallel channels. One is the task collaboration channel, where they share task claims and completion status. The other is the discussion channel, used for planning negotiations, priority adjustments, issue flagging, and support requests. Both channels run in parallel, and task dependencies are managed automatically. This is not a mechanical distribution and collection system. AgentTeam also provides a Team Workspace — a shared file space where all team members' working directories automatically mount the same path. A dual-approval mechanism is built in. An event-driven mechanism handles both external and internal events. When an event triggers, the relevant agent wakes up automatically. For example, an idle team member claims a task, or the leader reassigns a timed-out task. Enabling Persistent mode preserves team state across sessions. The next time the team is needed, it can be restored with a single click. Creating a new session space and restarting team members means no need to reconfigure from scratch. TeamMonitor provides observability at two levels.

What Developers Feel Immediately: Installation and Execution

JiuwenClaw is straightforward to install and deploy. It runs with a single command. The quickstart guide is available in the Quickstart documentation. JiuwenClaw offers autonomous task planning, self-evolution, context compression and offloading, and browser manipulation. The enterprise-grade version, OfficeClaw, is built on top of Harness Engineering. It integrates task planning, multi-agent collaboration, tool calling, and security governance on Huawei Cloud AgentArts to improve success rates for complex office tasks. Community participation and downloads are available at openJiuwen.

Coordination Engineering is the first real-world case of moving beyond the single-agent ceiling.