Next Leap into Binding Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’

How to make multiple agents work together as a special team – automatically dividing tasks, communicating effectively, and collaborating seamlessly?

The openJiuwen community has released the latest version of JiuwenClaw, which adds support for AgentTeam — the ability to collaborate with multiple agents. It suggests that the next leap beyond Harness Engineering be Coordinating Engineering.

In intensive tests, this team collaboration method has shown remarkable stability—team members have clear roles, work independently with seamless communication, and all workflows require no human intervention.

How strong is it, really?

It can automatically assemble a team of “well-trained” agents – and with that team, it can produce a solid, solid 200-page PPT in less than 20 minutes.

Project links:

Examining JiuwenClaw “Link Engineering” Works

Want deep insights without lifting a finger? A 200-page, content-rich PPT in less than 20 minutes.

In our test, we asked it to do an in-depth investigation of OpenClaw’s technology and break it down into 10 key features. For each feature, assign a dedicated agent to handle it. Each agent was responsible for producing 20 PPT slides, all under a unified theme. Finally, the 10 sets of slides were combined into a complete, 200-page technical presentation.

The whole process took less than 20 minutes. The resulting PPT was detailed, logically organized, and surprisingly efficient.

Technological Disparity: The Three Core Powers of the JiuwenClaw AgentTeam

AgentTeam’s core design philosophy is straightforward: it mimics how real-world teams collaborate.

• The Lead Agent is responsible for analyzing needs, building a team, and planning tasks.
• Collaborative Agents search for tasks, execute them independently, report results, and collaborate through a shared workspace.
• During execution, critical steps require the Leader’s approval, and error detection is automatic.

1. Hierarchical Independent Interaction: Leader Plays Smart, Teammates Play Automatically

JiuwenClaw AgentTeam delegates this responsibility to the Lead Agent himself.

What the Leader does:

• Builds a team dynamically: Assigns roles and members flexibly based on mission. If additional hands are needed during the execution, it can add or remove members later.
• Program activities: Breaks the goal down into concrete activities, establishing interdependencies (eg, “analysis can only begin after data collection is completed”).
• It also provides monitors: After creating tasks, it tracks progress in real time—who wants what, who completed what, who encountered problems—and adjusts accordingly.

What teammates do:

• Search for jobs proactively: Browse the job board and search for jobs that match their skills.
• Act independently: Complete their work at their workplace.
• Report results: Review the situation and inform the Leader and other dependents.

Team members drive core workflows through task collaboration—seek, do, finish, unlock downstream tasks—discuss plans, discuss priorities, flag issues, request support.

Both channels work in parallel, depending on the tasks being managed automatically—not just machine distribution and assembly.

2. Team Workspace: Shared Team File Space

JiuwenClaw AgentTeam solves this with Team Workspace—a team-level shared file space that all members can transparently access. Each group’s working directory automatically includes a shared path that points to the same group’s workspace.

3. Full Lifecycle Management: From Plan Approval to Automated Error Detection

3.1 Leader’s consent

AgentTeam provides a two-layer authentication method:

• Editing mode: For important tasks, the Team Partner first submits the execution plan for the Leader’s approval.
• Tool approval: If a Partner needs to perform sensitive work (eg, delete files, call external APIs, modify shared settings), a leader’s permission is required.

3.2 Event‑Driven Mechanism

AgentTeam mitigates this with an event-driven mechanism, using both external and internal events:

• External events: Work status changes, member lifecycle changes, messages between members—any meaningful change triggers an event.
• Internal events: Self-testing events generated by the framework (mailbox polling, job board polling) act as a safety net.

After an event is triggered, the appropriate agents are automatically activated (eg, tasks to claim idle Teammates, Leader reassigns expired tasks)

3.3 Persisting Parties

When Persistent Mode is enabled, groups can be saved across sessions: The next time you need a group, you can restore it with one click—you create a new session space, restart the group members, and you’re good to go, without having to rebuild the group from scratch.

3.4 TeamMonitor

TeamMonitor provides visibility into two categories:

• Query API: Check team information, member countries, job progress, and other statuses at any time.
• Event broadcast: Sign up for group events in real time. Task completion, member state changes, messages sent/received… all events can be consumed one by one with an asynchronous iterator. You can create dashboards, logging systems, or trigger external workflows from these events. All team performance measures are traceable and auditable.

Core Underpinning: OpenJiuwen AgentTeam Architecture

The main technical principles of AgentTeam can be summarized in three points:

1. Consistent collaboration through a shared task list: All members share the same dynamic task list. Each agent independently seeks and executes tasks based on the group’s mission, job descriptions, and capabilities—ensuring the consensus of environmental information.
2. Dual-drive model for messages and tasks: Members drive critical workflows through task switching, while continuously chatting and negotiating through a message channel outside of the task system—covering everything from formal execution to informal communication.
3. Role engineering and tools: RolePolicy defines the norms of behavior and decision boundaries of the Leader and Team Workers within the team. TeamTools gives team members specific communication skills. The role determines “what must be done,” and the tools determine “what can be done.”

About JiuwenClaw

JiuwenClaw is a “Claw” Agent developed over the open community of Jiuwen. It naturally supports multi-agent interaction and agent evolution. The core design philosophy is simple: Understand what you want, and evolve automatically.

Beyond AgentTeam, JiuwenClaw is also very easy to install and deploy – one command to get it up and running. For quickstart, see: /blob/develop/docs/en/Quickstart.md

In addition, JiuwenClaw offers several advantages in independent task scheduling, evolution, context compression and loading, browser manipulation, and overall “lobster-like” management:

• Automated task management: always ready when you’re ready : JiuwenClaw features a task scheduling mode, which is an AI to-do list. Users can interrupt, add, or modify tasks at any time.
• Your self‑Developing skills: It continuously records these performance errors and feedback, analyzes their origin, and generates targeted improvement suggestions. The update approval window then appears to the user – the entire update is your call.
• Content compression and loading : Effectively reduce costs by managing content length.
• Layered Memory: It achieves long-term storage and intelligent retrieval of conditions and performance tracking.
• To trick the browser: It automatically accesses profile information such as cookies and local cache, easily taking over browser space.

About OfficeClaw

The enterprise-level version, OfficeClaw, built on the basis of Harness engineering, seamlessly integrates task scheduling, multi-agent collaboration, tool invocation, and security management into Huawei Cloud AgentArts, improving the success rate of complex office operations.

Join the Community & Explore openJiuwen

openJiuwen Download Links

JiuwenClaw Download Links

• JiuwenClaw on GitHub:
• JiuwenClaw on AtomGit:
• AgentArts on Huawei Cloud:
• OfficeClaw on Huawei Cloud:/officeclaw.html

_{Note: Thanks to the OpenJiuwen team for resources, photos, video, and other information.}

Michal Sutter is a data science expert with a Master of Science in Data Science from the University of Padova. With a strong foundation in statistical analysis, machine learning, and data engineering, Michal excels at turning complex data sets into actionable insights.