An all-powerful intern holding the keys to the building
So far, you have a single AI agent. It does everything: web search, email, system commands. It has access to everything. That is convenient, but it is dangerous.
A booby-trapped email can inject instructions. A malicious web result can get code executed. And since the agent has access to everything — sudo, files, network — the damage is unlimited.
This guide creates 4 specialized agents, each in its own trust zone.
By Myrko Federico — 25 years of experience.
The 4 agents
-
Main — orchestrator. Receives your requests, delegates. It doesn't do the work, it coordinates
-
System — infrastructure ops. Sudo, Docker, maintenance. Elevated privileges, never any external content
-
Comm — email, messaging. Docker sandbox — a booby-trapped email stays confined to a disposable container
-
Web — search, browsing. Docker sandbox — untrusted web content never touches your system
What the guide covers
- Two trust zones — Operator vs Untrusted Content
-
exec-approvals.json — per-agent execution policy
-
SOUL.md, AGENTS.md, TOOLS.md, IDENTITY.md, HEARTBEAT.md — the complete identity of each agent
-
Docker sandbox — custom image, isolated network
-
Tool policy — the web agent cannot send email, the comm agent cannot use sudo
The most in-depth guide in the series (60 KB). Nothing comparable exists in French.
What you get
- The complete guide (PDF + Markdown)
- 4 appendices: OpenClaw Config, MCP Protocol, OpenClaw Commands, Linux Commands
Open Source Server Stack series. All options.
Format: PDF + Markdown — Time: 3-4 hours