|

Next Leap to Harness Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’

ByRicardo

How to make a number of brokers work collectively like an elite workforce — autonomously dividing duties, speaking effectively, and collaborating seamlessly?

The openJiuwen group launched the newest model of JiuwenClaw, which provides help for AgentTeam — a multi-agent collaborative functionality. It proposes that the following leap past Harness Engineering is Coordination Engineering.

In in-depth assessments, this workforce collaboration mechanism has demonstrated outstanding stability —workforce members have clear roles, collaborate autonomously with seamless coordination, and the complete workflow requires no human intervention.

How hardcore is it, actually?

It can autonomously assemble a “well-trained” workforce of brokers — and with that workforce, it may produce a strong, logically rigorous 200‑web page technical PPT in beneath 20 minutes.

Project hyperlinks: https://github.com/openJiuwen-ai/jiuwenclaw

Testing JiuwenClaw “Coordination Engineering” in Action

Want deep insights with out lifting a finger? A 200‑web page, content material‑wealthy PPT in beneath 20 minutes.

In our trial, we requested it to conduct an in‑depth investigation of OpenClaw know-how and break it down throughout 10 core points. For every facet, it assigned a devoted agent to take cost. Each agent was accountable for producing 20 PPT slides, all beneath a unified theme. Finally, the ten units of slides had been merged into a whole, 200‑web page technical presentation.

The complete course of took lower than 20 minutes. The ensuing PPT was detailed, logically structured, and impressively environment friendly.

Technical Breakdown: Three Core Capabilities of JiuwenClaw AgentTeam

The core design philosophy of AgentTeam is easy: simulate how real-world groups collaborate.

  • A Leader Agent is accountable for requirement evaluation, workforce constructing, and activity planning.
  • Multiple Teammate Agents declare duties, execute independently, report outcomes, and collaborate by means of a shared workspace.
  • During execution, key milestones require Leader approval, and fault restoration is computerized.

 1. Hierarchical Autonomous Collaboration: Leader Orchestrates Intelligently, Teammates Execute Autonomously

JiuwenClaw AgentTeam delegates this duty to the Leader Agent itself.

What the Leader does:

  • Dynamically builds the workforce: Assigns roles and members dynamically primarily based on the purpose. If extra arms are wanted mid-execution, it may add or take away members on the fly.
  • Plans duties: Breaks down the purpose into concrete duties, establishing dependencies (e.g., “evaluation can solely begin after information assortment is full”).
  • Assigns and displays: After creating duties, it tracks progress in actual time—who claimed what, who accomplished what, who bumped into points—and adjusts accordingly.

What Teammates do:

  • Claim duties proactively: Browse the duty board and declare duties that match their capabilities.
  • Execute independently: Complete their work inside their very own workspace. 
  • Report outcomes: Update the standing and notify the Leader and different dependents.

Team members drive the core workflow by means of activity collaboration—claiming, executing, finishing, unblocking downstream duties—discussing plans, negotiating priorities, flagging points, requesting help.

Both channels run in parallel, with activity dependencies managed robotically—not merely mechanical distribution and aggregation.

2. Team Workspace: A Shared Team File Space

JiuwenClaw AgentTeam solves this with Team Workspace—a workforce‑degree shared file house that every one members can transparently entry. Each Teammate’s working listing robotically mounts a shared path pointing to the identical workforce workspace.

3. Full Lifecycle Management: From Plan Approval to Automatic Fault Recovery

3.1 Leader Approval

AgentTeam gives a two‑layer approval mechanism:

  • Plan mode: For necessary duties, a Teammate first submits an execution plan for Leader approval. 
  • Tool approval: When a Teammate wants to carry out a delicate operation (e.g., deleting recordsdata, calling exterior APIs, modifying shared configurations), Leader approval is required.
3.2 Event‑Driven Mechanism

AgentTeam mitigates this with an occasion‑pushed mechanism, utilizing each exterior and inside occasions:

  • External occasions: Task state adjustments, member lifecycle adjustments, inter‑member messages—any significant change triggers an occasion.
  • Internal occasions: Framework‑generated self‑test occasions (mailbox polling, activity board polling) act as a security web.

After an occasion is triggered, the related brokers are robotically woke up (e.g., idle Teammates declare duties, the Leader reassigns timed-out duties)

3.3 Persistent Teams

With Persistent mode enabled, groups may be preserved throughout classes: The subsequent time you want the workforce, you may restore it with one click on—create a brand new session house, restart the workforce members, and also you’re prepared to go, with out rebuilding the workforce from scratch.

3.4 TeamMonitor

TeamMonitor offering observability in two dimensions:

  • Query API: Check workforce info, member states, activity progress, and different statuses at any time.
  • Event stream: Subscribe to workforce occasions in actual time. Task completions, member state adjustments, messages despatched/acquired… all occasions may be consumed one after the other by way of an asynchronous iterator. You can construct dashboards, logging programs, or set off exterior workflows from these occasions. Every step of the workforce’s operation is traceable and auditable.

Core Underpinning: openJiuwen AgentTeam Architecture

The core technical rules of AgentTeam may be summarized in three factors:

  1. Consistent collaboration by way of a shared activity listing: All members share the identical dynamic activity listing. Each agent independently claims and executes duties primarily based on the workforce purpose, activity definitions, and its personal capabilities—making certain pure info consistency.
  2. Dual‑drive mannequin of messages and duties: Members drive the core workflow by means of activity transitions, whereas additionally constantly discussing and negotiating by way of a message channel exterior the duty system—protecting all the things from structured execution to unstructured communication.
  3. Role and power engineering: RoleCoverage defines the behavioral norms and resolution boundaries of the Leader and Teammates inside the workforce. TeamTools endows workforce members with particular coordination capabilities. The position determines “what needs to be performed,” and the instruments decide “what may be performed.”

About JiuwenClaw

JiuwenClaw is a “Claw” Agent developed on high of the openJiuwen open‑supply group. It natively helps multi‑agent collaboration and agent self‑evolution. The core design philosophy is straightforward: Understand what you need, and evolve autonomously.

Beyond AgentTeam, JiuwenClaw can also be very simple to set up and deploy – a single command will get you up and working. For a fast begin, refer to: https://github.com/openJiuwen-ai/jiuwenclaw/blob/develop/docs/en/Quickstart.md

In addition, JiuwenClaw provides a number of benefits in autonomous activity planning, self‑evolution, context compression and offloading, browser manipulation, and total “lobster‑like” dealing with:

  • Autonomous activity administration: at all times prepared if you find yourself : JiuwenClaw incorporates a activity planning mode, which is actually a to‑do listing for the AI. Users can dynamically interrupt, append, or modify duties at any time. 
  • Self‑evolving Skills: Proactively information these execution errors and suggestions, analyzes the basis trigger, and generates focused enchancment options. An evolution approval window then pops up for the person – each replace is your name.
  • Context compression & offloading : Effectively reduces prices by managing context size.
  • Layered Memory: Achieves long-term storage and clever retrieval of situations and operation traces. 
  • Browser manipulation: Automatically accesses profile info like cookies and native cache, seamlessly taking up the browser atmosphere. 

About WorkplaceClaw

The enterprise-grade model, WorkplaceClaw, constructed on the Harness engineering basis, seamlessly integrates activity planning, multi-agent collaboration, software invocation, and safety governance on Huawei Cloud AgentArts, bettering the success charge of advanced workplace duties.

Join the Community & Explore openJiuwen

openJiuwen Download Links

JiuwenClaw Download Links

  • JiuwenClaw on GitHub: https://github.com/openJiuwen-ai/jiuwenclaw
  • JiuwenClaw on AtomGit: https://gitcode.com/openJiuwen/jiuwenclaw
  • AgentArts on Huawei Cloud：https://www.huaweicloud.com/product/agentarts
  • WorkplaceClaw on Huawei Cloud：https://www.huaweicloud.com/product/agentarts/officeclaw.html

Note: Thanks to the OpenJiuwen workforce for the assets, photographs, video, and different particulars.

The put up Next Leap to Harness Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’ appeared first on MarkTechPost.

Similar Posts