Agent Collaboration Protocol#

A complete protocol layer for building structured, multi-agent systems.

Provides the rules, templates, and operational procedures to ensure agents work together consistently.

What's Included#

ComponentDescriptionAGENT_COLLAB_PROTOCOL.mdCore protocol: L0/L1/L2 hierarchy, JSON task contracts, <agent-response> format, error handling, spawn vs send patternsOPERATIONS.mdOperational guide: how to add/remove domains, file ownership matrix, protocol change propagationNEW_AGENT_CHECKLIST.md8-step mandatory checklist for creating new sub-agents (no skip, no placeholder)TEMPLATE_AGENTS.mdStandard AGENTS.md template with protocol implementation tableTEMPLATE_IDENTITY.mdIdentity card template (Agent ID, layer, upstream/downstream, model)TEMPLATE_SOUL.mdValues and red lines template (core belief, values, anti-patterns)TEMPLATE_USER.mdService object template (upstream, downstream consumers, service standards)scaffold-domain.shShell script to auto-generate a new domain workspace from templates

Quick Start#

Install from ClawHub#

bash

npx clawhub install agent-collab-protocol

Or use locally#

bash

# Protocol files are in: skill/agent-collab-protocol/references/
# Copy to your protocol directory:
cp -r skill/agent-collab-protocol/references/ ~/.openclaw/protocol/

Create a new business domain#

bash

cd ~/.openclaw/protocol
bash scaffold-domain.sh fintech 金融产品经理 "fintech-analyzer:金融分析师"

Architecture Overview#

text

Layer 0 (Orchestrator)          Identify domain → dispatch to Manager → verify
    ↓ (natural language)
Layer 1 (Manager)               Understand → decompose → dispatch → aggregate
    ↓ (JSON Task Package)
Layer 2 (Specialist)            Execute → deliver via <agent-response>

Key Protocol Rules#

• L0 dispatches to L1 Managers via natural language
• L1 decomposes tasks into JSON Task packages for L2 Specialists
• L2 executes and returns <agent-response> blocks
• Tasks that fail: retry 1 time → downgrade → report as partial/failed
• 3-layer maximum depth (L0 → L1 → L2)

Usage#

For Framework Authors#

Reference AGENT_COLLAB_PROTOCOL.md to implement protocol compliance in your agent runtime.

For Agent Developers#

1. Copy the relevant TEMPLATE_*.md files
2. Fill in your agent's business-specific content
3. Register in your domain configuration
4. Follow NEW_AGENT_CHECKLIST.md for new agents

For Team Leads#

Use NEW_AGENT_CHECKLIST.md as a mandatory gating process — each team member must check off all 8 steps before a new agent is considered complete.

Files#

text

agent-collab-protocol/
├── SKILL.md                       ← This file
├── _meta.json                     ← Package metadata
└── references/
    ├── AGENT_COLLAB_PROTOCOL.md   ← Core protocol (413 lines)
    ├── OPERATIONS.md              ← Operational guide
    ├── NEW_AGENT_CHECKLIST.md     ← Mandatory checklist (156 lines, 64 checks)
    ├── TEMPLATE_AGENTS.md         ← AGENTS.md template
    ├── TEMPLATE_IDENTITY.md       ← Identity card template
    ├── TEMPLATE_SOUL.md           ← Values & red lines template
    ├── TEMPLATE_USER.md           ← Service object template
    └── scaffold-domain.sh         ← Domain scaffolding script

Changelog#

VersionDateChanges1.0.02026-06-04Initial public release