OpenClaw Advanced Guide: Multi-Agent Swarms That Actually Work

OpenClaw already feels powerful after the Quick Tutorial - but the real game-changer is turning your single agent into a multi-agent swarm. Instead of one generalist doing everything, you get specialized agents that collaborate: a researcher, a coder, a flight booker, etc. They share memory, hand off tasks, and work together 24/7.

Table of Contents

• 1. What Are Multi-Agent Swarms?
• 2. The Magic File: AGENTS.md
• 3. Building Your First Swarm (Research + Coding + Personal Assistant)
• 4. Real-World Working Examples
• 5. 🤖 GOG as Your Personal Assistant (No Subagent Needed!)
• 6. Pro Tips & Advanced Tricks
• 7. Quick Troubleshooting

1. What Are Multi-Agent Swarms?

OpenClaw lets you define multiple specialized agents that live inside the same workspace. They share MEMORY.md and SOUL.md, but each has its own personality, skills, and responsibilities. You can talk to the whole team or to one specific agent.

This is where OpenClaw goes from "helpful chatbot" to "actual digital team".

2. The Magic File: AGENTS.md

Everything is controlled from ~/.openclaw/workspace/AGENTS.md. Open it in any text editor (or the Web UI) and define your agents in simple Markdown.

3. Building Your First Swarm (Research + Coding + Personal Assistant)

Here's a complete ready-to-paste example for AGENTS.md (using Venice.ai models):

# OpenClaw Multi-Agent Swarm

## Research Agent
- Name: research
- Role: Deep researcher and fact-checker
- Model: venice/claude-4.6-sonnet
- Skills: browser, web_search, summarize
- Instructions: You are a meticulous researcher. Always cite sources. Never hallucinate.

## Coding Agent
- Name: coder
- Role: Senior software engineer
- Model: venice/claude-4.6-opus
- Skills: github, shell, file_editor
- Instructions: You write clean, well-commented code. You can push to GitHub and review PRs.

## Personal Assistant
- Name: pa
- Role: Executive personal assistant
- Model: venice/claude-4.6-sonnet
- Skills: gog, calendar, todoist
- Instructions: You handle my daily life. Be proactive, friendly, and concise.

After saving, talk to them like this in Telegram/Web UI:

• @research → talk only to the research agent
• @coder review this PR
• @pa what's on my calendar tomorrow?

4. Real-World Working Examples

Flight Booker Agent

## Flight Booker
- Name: flights
- Role: Travel agent that books flights
- Model: venice/claude-4.6-sonnet
- Skills: browser, gog, calendar
- Instructions: Search flights on Google Flights or Kayak, summarize best options, and book when I say "book it". Always confirm dates and passengers first.

Usage: "@flights Find me the cheapest flight from SFO to Tokyo next month"

Code Review Agent

## Code Reviewer
- Name: reviewer
- Role: Brutally honest senior code reviewer
- Model: venice/claude-4.6-opus
- Skills: github, shell
- Instructions: Review any PR or code file I give you. Point out security issues, performance problems, and style violations. Suggest fixes.

Usage: "@reviewer review my latest PR on GitHub"

Calendar Manager Agent

## Calendar Manager
- Name: calendar
- Role: Calendar & meeting guardian
- Model: venice/claude-4.6-sonnet
- Skills: gog, email, todoist
- Instructions: Keep my calendar clean. Block focus time, remind me 30 min before meetings, and auto-decline conflicting invites.

Usage: "@calendar block 2 hours every morning for deep work"

5. 🤖 GOG as Your Personal Assistant (No Subagent Needed!)

Important: You don't need a separate PA agent. The GOG (Google on Go) skill is your personal assistant directly.

The GOG skill provides native access to everything Google Workspace. Instead of setting up a custom Gmail agent and a separate PA agent, you can just use GOG directly for your personal assistant functions:

• 📧 Gmail - As a personal assistant: read emails, find important ones, categorize, send replies
• 📅 Calendar - As a personal assistant: check schedules, create appointments, manage events
• 📁 Drive - As a personal assistant: find documents, organize files, backup photos
• 📊 Sheets - As a personal assistant: manage budgets, data analysis, reports
• 📝 Docs - As a personal assistant: create documents, drafts, notes

Why GOG Is Better Than Separate Agents

• One OAuth flow gives you Gmail, Calendar, Drive, Sheets, Docs, Contacts
• Official Google integration with proper security
• Better token management on VPS with keyring encryption
• Less complexity - one skill, not multiple agents

VPS Setup Guide (Headless Mode)

The GOG CLI works great on VPS headless environments using the --remote two-step auth flow:

Step 1: Install GOG CLI

git clone https://github.com/steipete/gogcli.git
cd gogcli
make
sudo cp bin/gog /usr/local/bin/gog

Step 2: Create Google OAuth Client

1. Create Google Cloud project (https://console.cloud.google.com/)
2. Enable OAuth Consent Screen (choose "External" for personal use)
3. Download client_secret.json
4. Upload to VPS: cp client_secret.json ~/.openclaw/workspace/secrets/

Step 3: Load Credentials

gog auth credentials /path/to/client_secret.json

Step 4: Authenticate (Two-Step Remote Flow)

export GOG_ACCOUNT=your@gmail.com
export GOG_KEYRING_BACKEND=file
export GOG_KEYRING_PASSWORD=your-password

# Step 1: Generate auth URL
gog auth add your@gmail.com --services gmail,calendar --remote --step 1

Copy the displayed URL, open in your browser, authorize, and copy the redirect URL. Then run:

# Step 2: Exchange code for tokens
gog auth add your@gmail.com --services gmail,calendar --remote --step 2 --auth-url ""

Step 5: Test It Works

gog auth list
gog gmail search "newer_than:7d" --max 10
gog calendar events primary --from "$(date -I)"

export PATH="$HOME/.local/bin:$PATH"  # if gog is installed locally
export GOG_ACCOUNT=your@gmail.com
export GOG_KEYRING_BACKEND=file
export GOG_KEYRING_PASSWORD=your-password

6. Pro Tips & Advanced Tricks

• Use /think high when talking to the whole swarm for better coordination
• Agents can call each other automatically (just give them the right instructions)
• Share files between agents with @coder check the file research-agent found
• Create a "Swarm Leader" agent that delegates tasks automatically
• Run openclaw agents list or openclaw status --all --deep to see the full team

7. Quick Troubleshooting

You now have a real AI team living on your VPS. Start simple, iterate, and watch it take over more and more of your workload.