OpenClaw Multiple Agents: 13 on One Mac Mini

Most people install OpenClaw and get one agent doing one thing. That's fine for personal use. But what if you want an entire team? A COO, a social media manager, a video editor, a newsletter writer, an analytics scraper, an intelligence crawler, and a community manager. All running on one Mac Mini. All coordinated. All working while you sleep.

The 13-Agent Team: Every AI Agent and Its Role

Every agent has a single job. No overlap. No confusion.

Agent	Role	What They Do
Marc	COO	Coordinates everything. Dispatches tasks. Makes decisions. Talks to me on Telegram.
Bob	Ops	Health checks every 2 hours. Monitors all crons. Fixes or escalates.
Dan	X Growth	Manages @profitfounder. 3 posts/day via Typefully API.
Jimmy	YouTube	Monitors for new episodes. Extracts clips. Maintains transcript archive.
Tyler	Newsletter	Writes drafts from transcripts. Pushes to Notion. Teasers only.
Claude	Copy Editor	Scans transcripts for clip moments. Writes posts. Quality scores 1-10.
Adrien	Video Editor	Whisper transcription, smart cuts, silence removal, subtitle burn-in.
Mona Lisa	Sponsorship	Researches targets. Personalized X DM outreach. Pipeline in Notion.
Loop	Analytics	Daily scrape of YT, X, IG, Skool, Beehiiv. Weekly correlation analysis.
Crawly	Intelligence	Daily web crawl: Reddit, HN, YouTube, X. READ ONLY.
Billy	Community	Scans Skool daily. Content gaps. Module drafts. Churn tracking. READ ONLY.
Ariane	Notion	Daily cleanup. Moves misplaced pages. Archives stale items.
Alfred	Personal	Real-life tasks. Contractors, errands, local services.

How OpenClaw Multi-Agent Coordination Works

You don't run 13 separate OpenClaw instances. You run one gateway with one main agent. (Need to set up the hardware first? See our Mac Mini setup guide or VPS setup guide.) (Marc). Marc spawns sub-agents when work needs to happen.

Florian (human)
  |
  Marc (main agent, COO)
  |
  +-- Bob (dispatcher)
  |     +-- monitors all other agents
  |     +-- health checks every 2 hours
  |
  +-- Content Pipeline
  |     +-- Jimmy (YouTube / clips)
  |     +-- Claude (copy editor)
  |     +-- Adrien (video editing)
  |     +-- Dan (X posting)
  |     +-- Tyler (newsletter)
  |
  +-- Growth & Revenue
  |     +-- Mona Lisa (sponsorship)
  |     +-- Loop (analytics)
  |     +-- Billy (Skool community)
  |
  +-- Operations
        +-- Crawly (intelligence)
        +-- Ariane (Notion)
        +-- Alfred (personal)

Why This Architecture Works

The tree structure is intentional. If every agent talked to every other agent, you'd get chaos. 13 agents with full cross-communication means 156 possible communication paths. That's unmanageable.

Instead, communication flows through Marc. He's the single point of coordination. Think of it like a company org chart. The CEO doesn't need every employee talking to each other directly. Information flows up and decisions flow down.

The three clusters (Content Pipeline, Growth and Revenue, Operations) are also deliberate. Agents within a cluster sometimes share data through files, but they never directly spawn each other. Only Marc spawns sub-agents. This prevents circular dependencies where Agent A spawns Agent B which spawns Agent A again.

How Agents Communicate Through Files

Agents don't send messages to each other. They write files. Claude (copy editor) writes clip selections to notion/clips-queue/. Adrien (video editor) reads from that folder. Dan (X growth) reads finished clips from output/clips/.

This file-based communication has three advantages:

1. Auditability. I can see exactly what each agent produced by looking at the files.
2. Resilience. If one agent crashes, the files remain. The next agent picks up where things left off.
3. No race conditions. Each agent processes files on its own cron schedule. No two agents touch the same file at the same time.

How to Spawn OpenClaw Sub-Agents

When Marc needs Jimmy to check for new episodes, he spawns a sub-agent session:

sessions_spawn({
  task: "Check YouTube for new podcast episodes...",
  model: "anthropic/claude-opus-4-6",
  runTimeoutSeconds: 600
})

The sub-agent runs in isolation, does its job, and reports back. Marc gets the result and decides what to do next.

OpenClaw Agent SOPs: How to Write Standard Operating Procedures

Every agent has a file in sops/[agent-name].md. Without SOPs, agents drift, forget rules, and overlap with each other.

Every SOP contains:

• Role: One sentence. What this agent does.
• Responsibilities: Specific tasks. No ambiguity.
• Access/Logins: Exactly which tools and credentials this agent can use.
• Hard Rules: Things this agent must never do.

Example:

# SOP: Adrien - Video Editor Agent

## Role
Processes raw clips into polished, subtitled videos.

## Responsibilities
- Whisper transcription of raw clips
- Smart cuts (filler word removal)
- Silence removal
- Clean subtitle burn-in
- Compress for Telegram delivery

## Hard Rules
- Source clips MUST be h264
- Subtitles: 4-5 words per line max
- Never clip from the intro

OpenClaw Agent Permission System: Who Can Access What

Agent	Can Write To	Cannot Touch
Dan	X, Typefully	Notion deletions, memory files
Jimmy	Transcripts folder	Any external platform
Crawly	Nothing (READ ONLY)	All platforms
Billy	Notion drafts only	Skool (never posts)
Tyler	Notion drafts only	Beehiiv send (I send)
Marc	Memory, coordination	Social media, email

The Full 24-Hour OpenClaw Cron Schedule

Time	Agent	Task
12:30 AM	Crawly	Intelligence crawl (Reddit, HN, YouTube, X)
12:45 AM	Crawly	Team brief → Telegram
1:00 AM	Jimmy	Night research + episode check
2:00 AM	Claude	Scan transcripts for clips
4:00 AM	Bob	Clip QA gate
5:50 AM	Loop	Usage cost tracker
6:00 AM	Ariane	Notion cleanup + War Room
6:30 AM	Loop	Daily analytics scrape
7:00 AM	Marc	Morning brief → Telegram
8:00 AM	Billy	Skool community scan
9:00 AM	Billy	Churn monitor
1:00 PM	Jimmy	YouTube research
7:00 PM	Bob	War Room
7:30 PM	Marc	Evening brief
11:00 PM	Ariane	Daily backup + Notion sync

Plus recurring: Git auto-backup every 30 min. Billy member count every 4h. Bob health check every 6h.

OpenClaw Shared Memory Architecture for Multiple Agents

13 agents need to share context without stepping on each other.

File	Purpose	Who Writes
MEMORY.md	Long-term curated memory	Marc only
memory/YYYY-MM-DD.md	Daily notes / raw logs	Marc, Bob
memory/regressions.md	Failure guardrails	Any agent
memory/friction-log.md	Instruction contradictions	Any agent
memory/predictions.md	Decision calibration	Any agent
memory/context-holds.md	Temporary priorities with expiry	Marc

Master Rules: The Constitution for All OpenClaw Agents

On top of individual SOPs, master-rules.md applies to ALL agents:

• Every deliverable goes to Notion. Non-negotiable.
• No fake data. No sample content. Make it real.
• Write it down immediately. Rules go to memory that same turn.
• 80/20 only. If it doesn't move the goals, skip it.
• Model strategy: Opus 4.6 everywhere. No exceptions.
• All outreach on X DMs only. No cold emails. No bulk.

Resource Management: Running 13 Agents Efficiently

Running 13 agents on one machine sounds resource-intensive. It's not. Here's why.

OpenClaw agents don't run simultaneously. They run on cron schedules. At 1 AM, Jimmy runs for 5 minutes. At 2 AM, Claude runs for 10 minutes. At 4 AM, Bob runs for 3 minutes. They take turns. The Mac Mini is idle 80% of the time.

The actual resource bottleneck is API concurrency, not hardware. If two agents fire at the same time and both make API calls, you might hit Anthropic's rate limit. The solution: stagger cron schedules by at least 15 minutes.

Model Assignment by Task Complexity

Not every agent needs the most expensive model. Here's how I assign models to optimize cost:

Agent	Model	Why
Marc (COO)	Claude Opus 4.6	Needs best reasoning for coordination
Claude (Copy Editor)	Claude Opus 4.6	Writing quality matters
Bob (Ops)	Claude Opus 4.6	QA decisions require nuance
Dan (X Growth)	Claude Opus 4.6	Post quality directly affects metrics
Loop (Analytics)	Claude Opus 4.6	Data analysis needs accuracy
Crawly (Intel)	Claude Opus 4.6	Research summarization
Ariane (Notion)	Claude Opus 4.6	Organization tasks

In theory, you could use Sonnet or Haiku for simpler agents like Ariane (Notion cleanup) or Loop (data scraping). In practice, I run Opus everywhere because the quality difference matters and $150/month is cheap compared to hiring anyone.

Scaling: When to Add More Agents

Don't add agents for the sake of having more agents. Add them when you notice one agent doing two unrelated jobs. My rule: if an agent's SOP has two sections that could operate independently, split it into two agents.

Example: Marc used to handle both coordination AND Skool community management. The community work was distracting from his coordination job. So I split Billy out as a dedicated community agent. Marc's performance immediately improved because his context was focused.

What Running 13 AI Agents Actually Costs

Item	Cost
Mac Mini M4	$700 one-time
Claude Opus API	~$100-150/month
Electricity	~$3/month
Total monthly	Under $155/month

Compare that to hiring even one freelancer.

Lessons After Running 13 OpenClaw Agents for 2 Months

1. Start with one agent. Add slowly. I started with Marc, added Bob, then specialists as I understood the workflows.
2. SOPs prevent chaos. Write the SOP before spawning the agent.
3. Permissions matter. Build the access matrix before something gets accidentally deleted.
4. The dispatcher pattern is key. Marc → Bob → everyone else. Without this layer, the main agent gets overwhelmed.
5. Memory files are your continuity. If it's not written down, it doesn't exist.
6. Crons are the real power. Having agents work on a schedule without you asking is transformative.

Get the Complete Multi-Agent Template Pack

Inside OpenClaw Lab, members get the complete template pack: every SOP, the access matrix, the cron schedule, the memory architecture, and the coordination playbook. Copy-paste it, swap in your business context, and you have a 13-agent team running in an afternoon.

New to OpenClaw? Start with installopenclawnow.com to get your first agent running.

Frequently Asked Questions

How do I set up multiple AI agents in OpenClaw?

Create separate agent workspaces, each with its own SOUL.md personality, AGENTS.md instructions, and memory files. Configure agent routing in the gateway to direct different contacts or channels to different agents. Each agent runs in isolation.

How many AI agents can OpenClaw run simultaneously?

OpenClaw can run 13 or more agents simultaneously on a single Mac Mini. Each agent has its own workspace and cron schedule. The limiting factor is your LLM API budget, not hardware or software constraints.

Can OpenClaw agents communicate with each other?

Yes, OpenClaw agents can communicate through shared workspace files, sub-agent spawning, and dispatcher patterns. A coordinator agent can assign tasks to specialized agents and receive their results. This enables complex multi-agent workflows.

What is the best way to organize multiple OpenClaw agents?

Give each agent a clear, narrow responsibility with its own SOP file. Use a dispatcher agent to coordinate tasks. Keep agents isolated in separate workspaces to prevent conflicts. Name agents clearly so you know what each one handles.

Do multiple OpenClaw agents share the same API key?

Yes, multiple agents can share the same LLM API key. All API usage is billed to one account. You can also assign different API keys or different models to different agents to optimize costs based on task complexity.