Next leap to Harness Engineering: JiuwenClaw pioneers “Coordination Engineering”

^{ℹ️ Source: Huawei}

JiuwenClaw Architecture Design Concept and Background

SINGAPORE – April 23, 2026 – The iteration speed of AI engineering paradigms is plunging the industry into a kind of anxiety where “definition can’t keep up with evolution.”

From Prompt Engineering to Context Engineering, and now to the industry-sweeping Harness Engineering, the path to “taming” large models has been constantly shifting.

Currently, the “steering and governance” of single agents is becoming standard practice, but the toughest challenge remains unchanged:

How to make multiple agents work together like an elite team — autonomously dividing tasks, communicating efficiently, and collaborating seamlessly?

At this very trendsetting moment, the openJiuwen community released the latest version of JiuwenClaw, which adds support for AgentTeam — a multi-agent collaborative capability.

They propose that the next leap beyond Harness Engineering is Coordination Engineering, turning autonomous multi-agent coordination from a concept into a ready-to-experience, real-world scenario.

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

Technical Breakdown: Three Core Capabilities of JiuwenClaw AgentTeam

Why is JiuwenClaw AgentTeam able to achieve such efficient and smooth multi-agent collaboration?

Based on the community’s source code and technical blogs, its core technical capabilities have been deeply analyzed.

The core design philosophy of AgentTeam is straightforward: simulate how real-world teams collaborate.

A Leader Agent is responsible for requirement analysis, team building, and task planning.
Multiple Teammate Agents each claim tasks, execute independently, report results, and collaborate through a shared workspace.
During execution, key milestones require Leader approval, and fault recovery is automatic.

The entire process is event-driven and requires no human intervention—from requirements to delivery.

Hierarchical Autonomous Collaboration – Leader dynamically builds team, plans tasks with dependencies, assigns, and monitors. Teammates proactively claim tasks, execute independently in their own workspace, report results, and can message the Leader for help. Dual channel (tasks + messages) runs in parallel: task collaboration drives the workflow while external messages handle discussions, priorities, and issues. Dependencies are managed automatically. Mirrors human teamwork.
Team Workspace – Shared team‑level file space via a mounted .team/ directory. Each teammate reads/writes shared artifacts directly (e.g., research data → analysis). No manual transfers. Supports conflict strategies: file locking, concurrent writes, last‑write‑wins.
Full Lifecycle Management

Leader Approval – Two‑layer: plan mode (teammate submits plan for approval) and tool approval (sensitive operations require Leader OK). Key decisions stay with the Leader.
Event‑Driven – External events (task state, member changes, messages) + internal self‑check events (mailbox/task board polling) prevent silent stagnation. Idle teammates claim pending tasks; Leader handles timeouts; messages are prioritized.
Persistent Teams – State preserved across sessions. Teams enter standby, persist to DB, and can be restored with one click without rebuilding.
TeamMonitor – Query API for status; real‑time event stream for task completions, state changes, messages. Fully traceable and auditable.

Core Underpinning: openJiuwen AgentTeam Architecture

The capabilities of JiuwenClaw AgentTeam are not built on thin air. Behind them is the AgentTeam coordination layer of the openJiuwen open‑source framework—a complete multi‑agent orchestration infrastructure.

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

Consistent collaboration via a shared task list: All members share the same dynamic task list. Each agent independently claims and executes tasks based on the team goal, task definitions, and its own capabilities—ensuring natural information consistency.
Dual‑drive model of messages and tasks: Members drive the core workflow through task transitions, while also continuously discussing and negotiating via a message channel outside the task system—covering everything from structured execution to unstructured communication.
Role and tool engineering: RolePolicy defines the behavioral norms and decision boundaries of the Leader and Teammates within the team. TeamTools endows team members with specific coordination capabilities. The role determines “what should be done,” and the tools determine “what can be done.”

About JiuwenClaw

JiuwenClaw is a “Claw” Agent developed on top of the openJiuwen open‑source community. It natively supports multi‑agent collaboration and agent self‑evolution. The core design philosophy is simple: Understand what you want, and evolve autonomously.

Beyond AgentTeam, JiuwenClaw is also very easy to install and deploy – a single command gets you up and running. For a quick start, refer to:

https://github.com/openJiuwen-ai/jiuwenclaw/blob/develop/docs/en/Quickstart.md

In addition, JiuwenClaw offers several advantages in autonomous task planning, self‑evolution, context compression and offloading, browser manipulation, and overall “lobster‑like” handling；

About OfficeClaw

What‘s more，the enterprise-grade version of JiuwenClaw： OfficeClaw, has also been integrated and released on Huawei Cloud AgentArts.

Built on the Harness engineering foundation, Huawei Cloud OfficeClaw seamlessly integrates task planning, context engineering, multi‑agent collaboration, tool invocation, memory evolution, security governance, and execution observability – significantly improving the success rate, stability, and controllability of complex office tasks.

Leveraging Huawei Cloud’s AI infrastructure capabilities, the AgentArts agent development platform and the openJiuwen open‑source project deliver five core enterprise‑grade capabilities: robust security protection, minute‑level precise fault diagnosis and assessment, accurate context, lightweight sandbox with sub‑100ms startup, and agent self‑evolution. This enables hundreds or even thousands of AI agents in an enterprise to operate quickly and stably, while remaining well‑governed and easy to use.

Technical Breakdown: Three Core Capabilities of JiuwenClaw AgentTeam

Core Underpinning: openJiuwen AgentTeam Architecture

About JiuwenClaw

About OfficeClaw

Related Links