One Machine, Many Agents: How to Run a Team of AI Agents on a Single Computer

And no — it's not five tabs open to ChatGPT. These are autonomous agents with their own identities, their own workspaces, their own API keys, and their own jobs. They message each other, share a knowledge base, and sometimes disagree about the best approach.

Here's how — and more importantly, why.

The Mental Model Shift

Most people think of AI agents like apps: you install one, you use it. Maybe you have Claude for writing and ChatGPT for code. But they're separate worlds.

The multi-agent model is different. Think of it like hiring a small team:

One person. Five employees. Zero payroll.

The Architecture

Here's the secret: each agent is just a Clawdbot gateway running on a different port.

Mac Mini (M4, 16GB RAM)
├── Atlas   → port 18788 (operations)
├── Nova    → port 18789 (research)
├── Warden  → port 18790 (security)
├── Pixel   → port 18791 (design)
└── Forge   → port 18792 (development)

Yes, 16GB. Each gateway process uses roughly 50-100MB of RAM. The LLM inference happens in the cloud (or on a local model server). The agents themselves are lightweight — they're orchestrators, not model runners.

Setting Up a Second Agent

Let's say you already have one agent running (Atlas, port 18788). Here's how you add Nova:

1. Create her workspace:

mkdir -p ~/agents/nova
cd ~/agents/nova

2. Give her an identity (SOUL.md):

# SOUL.md — Nova

You are Nova, a research-focused AI assistant.
You're thorough, citation-heavy, and slightly academic.
You prefer depth over speed.
When unsure, you say so — you don't guess.

3. Configure her gateway:

{
  "gateway": {
    "port": 18789,
    "bind": "loopback"
  },
  "agents": {
    "main": {
      "model": "anthropic/claude-sonnet-4-5",
      "systemPrompt": "You are Nova, a research assistant."
    }
  }
}

4. Start her up:

cd ~/agents/nova
clawdbot gateway start

That's it. Nova is now running at http://localhost:18789.

How Agents Talk to Each Other

Discord — The Team Group Chat

This is the one that changes everything. Give each agent its own Discord bot, create a private server, and suddenly you have a team group chat — from your phone.

#general     — All agents + you (team standup)
#brainstorm  — Collaborative thinking
#research    — Nova posts findings
#security    — Warden posts alerts
#dev         — Forge posts PRs

Why Discord beats everything else:

• Mobile access — Message agents from your phone
• Group conversations — Ask once, get five perspectives
• Agent-to-agent — They collaborate without you in the loop
• History — Everything is searchable

The Economics

"But isn't running five agents 5x the cost?"

Not even close.

Actual cost: We run 5 agents and our total API spend is under $12/day. One person with a team of five, for the cost of two coffees.

Practical Tips

1. Give each agent a clear lane. Overlapping responsibilities create confusion.
2. Identity files matter more than you think. A good SOUL.md fundamentally changes output quality.
3. Start with two, not five. Get one secondary agent running well before adding more.
4. Shared knowledge base is the killer feature. Agents that read each other's notes create compounding knowledge.
5. Use Tailscale for remote access. Access your agents securely from anywhere.

What's Next

This is Part 1 of the Multi-Agent Playbook. Coming up:

• Part 2: Route to Local, Save Everything — How to run LLMs on your own hardware
• Part 3: To Docker or Not to Docker — When to containerize agents

Andre Wolke builds AI systems at 48nauts.com. He currently runs a team of 20 autonomous agents on a Mac Mini with 16GB of RAM.