From Solo Agents to Elite Teams: openJiuwen’s Coordination Engineering Enables Self‑Evolving AI Collaboration

Coordination Engineering is presented as the next step beyond Harness Engineering, addressing the need for multi‑agent collaboration in complex tasks such as full‑house renovation, large technical reports, and medical consultations. It defines a systematic stack—team orchestration, task scheduling, communication protocols, isolation mechanisms, fault recovery, and observability—to transform isolated agents into a coordinated team.

Agent Team Engine

The core of Coordination Engineering is the JiuwenClaw Agent Team Engine, which mimics real‑world team dynamics: a Leader Agent performs demand analysis, team formation, and task planning, while multiple Teammate Agents independently claim tasks, execute, and report results through a shared workspace. Leader‑driven global orchestration eliminates the need for manual sequencing, and a dual event‑driven mechanism prevents deadlocks by allowing idle teammates to pick up pending tasks and enabling the Leader to re‑plan or replace stalled agents.

Team Skills

Team Skills capture the entire collaboration workflow—roles, responsibilities, execution order, constraints, dependencies, and assets—into a simple directory structure:

<team-skill-name>/
├── SKILL.md          ← team purpose and members
├── roles/
│   ├── <role-a>.md
│   └── <role-b>.md
├── workflow.md       ← coordination sequence
├── bind.md           ← error handling and boundaries
├── dependencies.yaml ← external tool dependencies
└── examples/ | templates/ | assets/ …

Team Skills can be generated automatically by the "teamskill‑creator" expert: a natural‑language description is enough to produce a complete skill package, which can also convert existing single‑agent skills or be refined with additional roles.

Practical Demonstrations

Examples include assembling a renovation team (hard‑design, soft‑design, artist), creating a 23‑member medical expert consultation team that dynamically selects specialists based on user symptoms, and building a travel‑planning team that self‑evolves by adding a copy‑writing expert after detecting role overlap.

Team Skills Hub and Self‑Evolution

The Team Skills Hub platform enables uploading, searching, and downloading reusable skills across eight domains (programming, office productivity, content creation, multimodal media, data & research, compliance & law, lifestyle & health, finance). Skills evolve through a dual‑layer self‑evolution mechanism: the team‑level skill adapts by adding roles or refining workflows, while individual member skills automatically incorporate resolved errors and performance improvements. Evolutionary patches are stored as independent experience entries, preserving the original skill files and allowing quality‑scored review.

Open Source Availability

All components—Agent Team Engine, Team Skills, and the teamskill‑creator—are open‑source on AtomGit and GitHub (e.g., https://github.com/openJiuwen-ai/jiuwenclaw). The latest version supports cluster mode in Feishu channels and can be switched to planning or performance modes via simple commands.