A comprehensive design framework for orchestrating multiple AI agents through structured delegation, task management, and concurrency control.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install eh-multi-agent
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install eh-multi-agent using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Multi-Agent Coordination?
The Multi-Agent Coordination skill provides the architectural blueprints and protocols necessary for managing complex AI operations where multiple agents must collaborate seamlessly. By leveraging this Openclaw Skills resource, developers can transition from single-agent execution to sophisticated multi-agent systems that handle large-scale refactoring, parallel testing, and multi-phase research projects without synchronization errors.
At its core, the skill defines three primary delegation modes—Coordinator, Fork, and Swarm—each optimized for different context and concurrency requirements. It enforces technical rigors such as workspace isolation using git worktrees and file-level atomic locking to ensure that agents do not overwrite each other's work, providing a stable environment for high-output AI automation.
Multi-Agent Coordination Use Cases
- Orchestrating multi-phase tasks that require separate stages for research, synthesis, and implementation.
- Scaling homogeneous tasks like API endpoint testing across multiple parallel worker agents.
- Managing cross-module refactoring where file conflicts and dependency chains must be strictly controlled.
- Implementing blind review processes where a secondary agent verifies code without bias from the original implementation logic.
How Multi-Agent Coordination Works
- Mode Selection: Determine the optimal delegation pattern (Coordinator for complex phases, Fork for context inheritance, or Swarm for flat task lists).
- Task Initialization: Create a shared .coordination/tasks.json to track the lifecycle of sub-tasks from pending to done.
- Workspace Isolation: Spawn agents in dedicated git worktrees to prevent index conflicts and provide clean execution environments.
- Concurrency Control: Utilize atomic directory-based locks in the .claims/ directory to manage file ownership during write operations.
- Synthesis Gating: Enforce a mandatory synthesis phase where the coordinator digests worker outputs into a structured implementation plan.
- Verification: Route the final output through a separate review session to validate quality before merging results.
Multi-Agent Coordination Setup
To implement these coordination patterns within your Openclaw Skills environment, initialize the required directory structure and worker environments:
# Create coordination and claim tracking directories
mkdir -p .coordination .claims
# Add a dedicated worktree for a specific worker agent
git worktree add .worktrees/worker-01 -b work/worker-01 HEAD
# Initialize a task board (optional but recommended)
echo '{"tasks": [], "updated_at": "'$(date -u +%Y-%m-%dT%H:%M:%SZ)'"}' > .coordination/tasks.json
Multi-Agent Coordination Data Schema & Taxonomy
| File/Path | Purpose | Data Format |
|---|---|---|
.coordination/tasks.json |
Central registry for task status and agent assignments. | JSON with fields: id, status, claimed_by, result. |
.coordination/synthesis.md |
Mandatory synthesis document containing conclusions and action plans. | Markdown with Conclusion and Action Plan sections. |
.claims/*.lock |
File-level mutex locks to prevent editing conflicts. | JSON with session_id and expires_at (TTL). |
.coordination/review-result.json |
Structured output from blind review agents. | JSON with approved (bool) and issues list. |
Multi-Agent Coordination Advanced Features
- Cache-safe Forking: Child agents inherit the parent's prompt cache to drastically reduce token costs and latency.
- Atomic POSIX Locking: Uses directory creation as a reliable atomic lock mechanism for task claiming.
- Synthesis Gate Validation: Scripted checks ensure the coordinator has actually synthesized findings rather than just forwarding raw data.
- TTL-Based Deadlock Prevention: Automatic lock expiration (default 10 mins) ensures crashed agents do not permanently block file access.
- Worktree-based Merge Strategies: Support for --no-ff merges or patch-based reviews to integrate multi-agent contributions safely.
SKILL.md
Loading
Related Openclaw Skills
MBUSA Dealer & Inventory Skill
A specialized AI agent skill for querying live Mercedes-Benz USA dealership information and real-time vehicle inventory.
Skill Discoverer
Skill Discoverer is an intelligent L0 entry-layer engine that helps users find, combine, and learn how to use AI agent capabilities using vector-based similarity and task analysis.
Ouroboros
A high-performance relay skill for managing automated AI workflows and status tracking via Openclaw Skills.
Hyper-Advance Skill Matrix (Executive Technical Leadership)
A comprehensive technical leadership framework designed for CTOs and VPs to scale engineering organizations, optimize high-scale architectures, and drive business impact.
Stock Analyst
A professional-grade AI financial analyst providing real-time technical and fundamental stock insights across global markets.
Novel Generator
An automated storytelling framework that uses structured memory and pacing formulas to generate consistent, high-engagement web novels.
Featured*
