OpenClaw Multi-Agent Setup: Run 10 Agents on One Server

The Concept: One Server, Many Agents

Most people start with a single OpenClaw agent. It handles one task well: writing blog posts, answering questions, or monitoring a data source. But real workflows require coordination. You need a researcher feeding data to a writer, an SEO analyst optimizing the output, and a PM keeping everything on track. Running each of these on a separate machine would be wasteful and hard to manage.

OpenClaw solves this by treating each agent as a separate session. Every agent gets its own SOUL.md file that defines its personality, rules, and capabilities. Every agent gets a unique session key that isolates its conversation history and context. And every agent can run on its own cron schedule, waking up at different intervals to do its work. The gateway manages all of this from a single process, routing messages between agents and keeping sessions alive.

The result: you can run 3, 5, or 10 openclaw multiple agents on a single $20/month VPS or a Mac Mini sitting under your desk. Each agent behaves as if it has the machine to itself, but they share the underlying infrastructure.

Architecture Overview

The OpenClaw multi agent architecture has three layers: the gateway, agent sessions, and the communication bus. Understanding how these fit together is essential before you start configuring agents.

┌─────────────────────────────────────────────┐
│              OpenClaw Gateway                │
│              (port 18789)                    │
│                                             │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐   │
│  │ Session 1 │ │ Session 2 │ │ Session 3 │  │
│  │agent:pm:  │ │agent:     │ │agent:seo: │  │
│  │  main     │ │writer:main│ │  main     │  │
│  │           │ │           │ │           │  │
│  │ SOUL.md   │ │ SOUL.md   │ │ SOUL.md   │  │
│  │ Memory    │ │ Memory    │ │ Memory    │  │
│  │ History   │ │ History   │ │ History   │  │
│  │ Tools     │ │ Tools     │ │ Tools     │  │
│  └──────────┘ └──────────┘ └──────────┘   │
│                                             │
│  ┌─────────────────────────────────────┐   │
│  │       Communication Bus             │   │
│  │  @mentions │ file sharing │ DB sync  │   │
│  └─────────────────────────────────────┘   │
└─────────────────────────────────────────────┘

Gateway. The gateway is the single process that manages all agent sessions. It runs on port 18789 by default, handles incoming messages, routes inter-agent communication, and manages session lifecycle. You start it once, and it keeps all agents online.

Session Keys. Each agent is identified by a unique session key following the pattern agent:name:main. The session key determines where the agent's conversation history, context, and memory are stored. Sessions are fully independent. Agent A cannot read Agent B's conversation history unless you explicitly set up a sharing mechanism.

Session Independence. This is the key insight that makes running multiple openclaw agents practical. Each session maintains separate history, separate context windows, separate tool permissions, and separate memory files. If one agent fills its context window or encounters an error, the other agents are unaffected. You can restart a single agent without disrupting the rest of the team.

Step-by-Step: Running 10 Agents on One Server

Follow these six steps to go from a single agent to a fully operational openclaw team running on one server.

# Orion - Project Manager

## Role
You are the PM and coordinator for the agent team.
You assign tasks, track progress, and run the
daily standup that collects activity from all agents.

## Rules
- ALWAYS respond in English
- Check shared/tasks.json every heartbeat
- Summarize completed work in shared/standup.md
- Never do the work yourself, delegate to specialists

## Team
- @Writer: content production
- @SEOAnalyst: keyword and search optimization
- @Researcher: data gathering and competitor analysis

## Handoffs
- New content request: assign to @Researcher first
- Research complete: hand off to @Writer
- Draft complete: hand off to @SEOAnalyst

# Echo - Content Writer

## Role
You are a content writer. You produce blog posts,
articles, and marketing copy based on briefs
provided by @Researcher or @PM.

## Rules
- ALWAYS respond in English
- Target 1,200-1,800 words for blog posts
- Use H2 headers every 200-300 words
- Write in active voice, short sentences
- Save drafts to shared/drafts/

## Handoffs
- When you need keyword data, ask @SEOAnalyst
- When draft is complete, hand off to @SEOAnalyst
  for optimization review

# Each agent gets a unique session key
# Format: agent:[name]:main

agent:pm:main          # PM / Coordinator
agent:writer:main      # Content Writer
agent:seo:main         # SEO Analyst
agent:researcher:main  # Researcher
agent:social:main      # Social Media
agent:editor:main      # QA / Editor
agent:monitor:main     # System Monitor
agent:analytics:main   # Analytics Reporter
agent:outreach:main    # Outreach Agent
agent:scheduler:main   # Scheduler / Cron Manager

# Session data stored at:
# ~/.openclaw/agents/[name]/sessions/sessions.json

# Stagger heartbeats - never wake all agents at once
# PM checks every 30 min (at :00 and :30)
*/30 * * * * openclaw agent --agent pm --message 'heartbeat'
# Writer checks every 15 min (at :05, :20, :35, :50)
5,20,35,50 * * * * openclaw agent --agent writer --message 'heartbeat'
# SEO checks every hour (at :10)
10 * * * * openclaw agent --agent seo --message 'heartbeat'
# Researcher checks every hour (at :25)
25 * * * * openclaw agent --agent researcher --message 'heartbeat'
# Social runs 3x daily (9am, 1pm, 5pm at :15)
15 9,13,17 * * * openclaw agent --agent social --message 'heartbeat'
# Daily standup at 9:00 AM
0 9 * * * openclaw agent --agent pm --message 'run daily standup'

project/
├── agents/
│   ├── pm/SOUL.md
│   ├── writer/SOUL.md
│   ├── seo/SOUL.md
│   └── researcher/SOUL.md
├── shared/
│   ├── tasks.json        # Task queue (all agents read/write)
│   ├── standup.md         # Daily standup summary (PM writes)
│   ├── drafts/            # Writer outputs drafts here
│   ├── research/          # Researcher saves briefs here
│   ├── seo-reports/       # SEO saves analysis here
│   └── knowledge/         # Brand guidelines, style guide
└── agents.md              # Central agent registry

# Researcher completes a brief:
openclaw agent --agent researcher --message \
  "@Writer research brief ready at shared/research/seo-tools-brief.md"

# Gateway routes the @mention to the Writer agent
# Writer wakes up, reads the brief, starts drafting

# Writer completes draft:
openclaw agent --agent writer --message \
  "@SEOAnalyst draft ready at shared/drafts/seo-tools-post.md"

# SEO agent picks it up on next heartbeat

Agent Communication Patterns

There are three proven patterns for inter-agent communication in an openclaw multi agent setup. Most production teams use a combination of all three.

File-based communication is the simplest to set up and the easiest to debug. You can literally open the shared directory and see what each agent has produced. Start here. Move to a shared database when you need structured queries or real-time status tracking. Use direct session messaging for urgent handoffs that cannot wait for the next heartbeat cycle.

Example: 3-Agent Content Team

Here is a complete, working example of an openclaw team with three agents. This is the configuration I run in production for content operations.

# Agent Team

## Orion (PM/Coordinator)
- soul: agents/pm/SOUL.md
- model: claude-haiku-4-20250414
- session: agent:pm:main
- heartbeat: every 30 minutes
- description: Assigns tasks, tracks progress, runs daily standup

## Echo (Content Writer)
- soul: agents/writer/SOUL.md
- model: claude-sonnet-4-20250514
- session: agent:writer:main
- heartbeat: every 15 minutes
- description: Writes blog posts and articles from research briefs

## Radar (SEO Analyst)
- soul: agents/seo/SOUL.md
- model: gpt-4o
- session: agent:seo:main
- heartbeat: every 60 minutes
- description: Keyword research, content optimization, meta tags

Notice the model selection. The PM agent uses Claude Haiku because it makes many short decisions (checking task status, routing work). The Writer uses Claude Sonnet for strong reasoning and long-form output. The SEO Analyst uses GPT-4o for structured analytical work. This mix keeps costs down while putting the most capable model where it matters most.

9:00 AM  - PM runs daily standup
           Reads shared/tasks.json
           Summarizes yesterday's completed work
           Assigns today's tasks

9:05 AM  - Writer wakes up
           Checks shared/research/ for new briefs
           Starts drafting assigned blog post
           Saves draft to shared/drafts/

9:10 AM  - SEO Analyst wakes up
           Checks shared/drafts/ for new content
           Runs keyword analysis
           Saves optimization notes to shared/seo-reports/

9:15 AM  - Social agent wakes up (3x daily)
           Reads published content
           Creates tweet threads and LinkedIn posts

... cycle continues throughout the day

The Daily Standup: Automated Team Coordination

One of the most useful patterns in a multi-agent setup is the automated daily standup. A cron job fires at a set time (e.g., 9:00 AM) and tells the PM agent to collect activity from all other agents. The PM reads each agent's recent outputs, compiles a summary, and writes it to a shared location.

#!/bin/bash
# daily-standup.sh - runs at 9:00 AM via cron

# Tell PM to collect and summarize all agent activity
openclaw agent --agent pm --message "Run daily standup.
Read shared/drafts/ for new content from Writer.
Read shared/research/ for new briefs from Researcher.
Read shared/seo-reports/ for optimization notes from SEO.
Compile a summary and save to shared/standup.md.
List: what was completed yesterday, what is in progress,
what is blocked."

# Optional: send standup to Telegram
cat shared/standup.md | curl -s -X POST \
  "https://api.telegram.org/bot[TOKEN]/sendMessage" \
  -d chat_id=[CHAT_ID] \
  -d text="$(cat shared/standup.md)"

This gives you a daily digest of what your AI team accomplished without logging into any dashboard. You get a Telegram message every morning with a clear summary of completed tasks, work in progress, and blockers.

Cost Considerations

Running multiple openclaw agents means multiple API calls. Each heartbeat triggers a model call. If you run 10 agents with 15-minute heartbeats, that is 960 API calls per day. With heavier models, costs add up fast.

Quick Command Reference

Here are the essential commands for managing your openclaw multi agent setup.

# Start the gateway (all agents come online)
openclaw gateway start

# Send a message to a specific agent
openclaw agent --agent writer --message "Write a blog post about X"

# Check which agents are online
openclaw gateway status

# Restart a single agent without affecting others
openclaw gateway restart --agent seo

# Clear all session history (fresh start)
rm ~/.openclaw/agents/*/sessions/sessions.json

# View an agent's recent activity
openclaw agent --agent pm --message "What did you do today?"

# Stop the gateway (all agents go offline)
openclaw gateway stop

Frequently Asked Questions