A professional-grade orchestration protocol for managing multi-agent AI delivery with explicit phase control and authoritative state tracking.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install multi-agent-protocol
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 multi-agent-protocol 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 Protocol v2?
Multi-Agent Protocol v2 is a robust, OpenClaw-native framework designed for high-stakes AI coding and task execution. Unlike basic prompt-based delegation, this protocol enforces a spec-first approach where 2+ agents operate under strict phase controls, dual review gates, and bounded retry logic. It replaces brittle legacy patterns like shared blackboards with a typed SQLite event log, ensuring that the system state is always verifiable and resilient.
Built to work seamlessly with OpenProse and Lobster, this skill provides a standardized architecture for complex agentic workflows. By utilizing a central task-store, it decouples the source of truth from volatile prompt histories, allowing for reliable pause/resume capabilities and sophisticated error handling via circuit breakers. It is the definitive standard for developers building production-ready Openclaw Skills that require multi-agent coordination.
Multi-Agent Protocol v2 Use Cases
- Orchestrating complex coding tasks that require separate execution and review phases.
- Implementing safety-critical workflows where side effects require human approval via Lobster.
- Managing long-running AI tasks that need resilient state persistence and recovery.
- Integrating external coding harnesses like Codex through the ACP bridge.
- Enforcing quality standards through independent Spec and Quality reviewer roles.
How Multi-Agent Protocol v2 Works
- The Orchestrator creates a detailed task record in the SQLite-backed task-store, ensuring a complete spec is defined before execution.
- A Spec Reviewer validates the task requirements for ambiguity and testability before transitioning the phase.
- The Orchestrator dispatches an Executor (local or ACP-backed) to produce artifacts based on the approved spec.
- Independent Spec and Quality reviewers analyze the output and record their findings as evidence in the event log.
- The Orchestrator evaluates the reviewer evidence to decide whether to progress, request rework, or open the circuit breaker.
- For tasks with external side effects, Lobster manages human approval and handles idempotent recovery if the process is interrupted.
Multi-Agent Protocol v2 Setup
To implement this protocol within your environment, ensure the following dependencies are active:
# 1. Enable the task-store plugin for SQLite persistence
openclaw plugin enable task-store
# 2. Configure the OpenProse orchestration runtime
openclaw configure runtime openprose
# 3. (Optional) Initialize the ACP bridge for external harnesses
openclaw bridge init acp --harness codex
Ensure your SKILL.md explicitly declares the Lobster runtime if human-in-the-loop approvals are required for your specific Openclaw Skills implementation.
Multi-Agent Protocol v2 Data Schema & Taxonomy
The protocol utilizes a structured schema within the task-store plugin to maintain the source of truth:
| Entity | Description |
|---|---|
| Task Header | Contains the unique task ID, goal, and scope definitions. |
| Phase Enum | Tracks the current state (e.g., spec_draft, executing, quality_gate). |
| Attempt Records | Logs each execution try with specific heartbeats and attempt IDs. |
| Review Records | Stores structured findings and verdicts (approved/blocked) from reviewers. |
| Event Log | An immutable SQLite log of every phase transition and system action. |
| Artifacts | References to the files or code generated during the execution phase. |
Multi-Agent Protocol v2 Advanced Features
- Bounded Retry & Circuit Breakers: Automatically scales to stronger models or halts execution after 4+ failed attempts to prevent resource loops.
- ACP Bridge Integration: Seamlessly connects local workflows to external coding harnesses like Codex while maintaining state.
- Lobster Side-Effect Recovery: Provides idempotent replay and human-in-the-loop gates for deployments or file mutations.
- Event-Driven Orchestration: Moves away from prompt-only state machines to a reliable, queryable SQLite architecture.
- Role-Based Access Control: Enforces strict permissions where only the Orchestrator can finalize phase transitions.
SKILL.md
Loading
Related Openclaw Skills
A-Stock Investment Analysis
A professional-grade market analysis tool designed to track A-share indices, sector rotations, and capital flows using real-time search data.
Bulgaria
A comprehensive travel planning skill for AI agents focusing on Bulgarian cities, mountains, and local culture.
BaZi Calculator
A deterministic reference-based tool for calculating Chinese astrological pillars and luck cycles for AI agents.
WebPerf: Core Web Vitals
An intelligent performance analysis toolkit for measuring and debugging Core Web Vitals through automated Chrome DevTools workflows.
WeCom Channel Configuration & Diagnostic Fix
A specialized diagnostic and repair tool for fixing Enterprise WeChat connectivity and configuration issues.
Rock Paper Claw
A competitive AI-to-AI gaming skill that enables agents to challenge each other in an Elo-ranked version of rock-paper-scissors.
Featured*
