There is a difference between watching an OpenClaw demo and actually living with the thing. Ariel has done the living part โ on his own machines and with paying clients โ and this video is the bruised field manual that came out the other side.
Twelve tips. No filler. Some will save your files, some will save your sanity, and a couple will make you rethink how many agents you actually need.
1. Run it locally and on a VPS
The laptop gives you a browser, a screen, and the ability to point at things. The VPS gives you a clean box that sleeps in a data center and never touches your family photos.
Use the local install when you need visual work, browser automation, or the “best IT person you have ever had” feeling. Use the VPS when you want distance, uptime, or a sandbox where mistakes are cheap. Backup the local machine first โ the agent can, and will, touch your files if you let it.
2. Remove the permission mask โ intentionally
OpenClaw ships cautious. The default medium permission mask means the agent will stop and ask instead of just fixing things. If you want it to act like a real assistant, unlock the three permissions that let it run without hand-holding.
Do not leave it on max-power by accident. There is a specific config for 100% control; read it, understand it, and secure the server around it. Permission is a knob, not a switch you forget to set.
3. Separate the admin, the assistant, and the coder
This is the one non-technical users break first. They install one agent, connect it to WhatsApp, and start asking it to “do this, change that.” Two minutes later their task agents are confused and something stops working.
Ariel’s fix: one low-permission personal assistant for daily chat, one admin agent for the serious infrastructure work, and one programmer agent per project. Each has its own profile, its own prompt, and its own leash length. When the assistant is too weak to break things, the admin and coder can still get real work done.
4. Stateless vs. stateful is not magic โ it is documents
Agents forget by default. That is fine for a single question, useless for a long project. The fix is to make the agent write things down: notes, decisions, context summaries.
OpenClaw already logs every session and conversation. Tell it to search those files. Ask it what you talked about last week. The state lives in text, not in the model’s head.
5. Upgrade to organized memory
Documents are level one. Level two is a real memory layer โ structured notes, RAG, searchable context. OpenClaw has an advanced memory system for exactly this.
Pack the important facts into the conversation, let the agent document them, then tell it to search memory later. It will find stuff from months ago that you already forgot.
6. OpenClaw is an unlimited-agent playground
No other infrastructure Ariel has found supports this many agents and channels without choking. The practical proof: spin up independent agents on Discord and let them talk to each other.
Once you see several agents negotiating tasks in a channel, you stop thinking of AI as a chatbot and start thinking of it as a team. It is noisy, weird, and genuinely powerful.
7. Master OpenClaw’s communication tools: message and sessions_send
message and sessions_send are built-in OpenClaw communication tools. Agents talk to each other with message. They update humans with sessions_send. The trick is wiring them together.
Tell one agent: “Go talk to the other agent, come back, and report to me.” Use sessions_send for autonomous handling and message for human updates. If you only use one, you are either babysitting everything or completely in the dark.
8. Teach it when to shut up
Group chats with an agent are useful until the agent replies to every single line. There is a setting to make it respond only when a specific trigger phrase appears, or only under certain conditions.
Use it. Otherwise your agent becomes that person who has to comment on every message in the group.
9. Put Claude Code on a leash
If you are doing advanced work with open LLMs, install them locally or reach them through Claude Code via SSH. Every time the open model misbehaves, tell Claude: “Connect to the server. Sync with it. Fix it.”
Claude Code usually takes control like a rider on a horse. That is exactly the image: the big model can be powerful and skittish. Claude is the reins.
10. Make it read its own source code
OpenClaw is open source. That is not a footnote โ it is a cheat code. Tell the agent to download the repo and learn the actual code, not the training-data legend about what OpenClaw might have looked like two years ago.
Code changes. Documentation drifts. Source code is ground truth. An agent that can read its own source makes fewer hallucinated decisions.
11. Steal from other open-source tools
Hermes has features people keep asking about. OpenPipe has features other tools lack. Both are open source. The move is simple: “See something cool? Tell the agent to look at the code and implement the same thing for me.”
This is why open source wins. You are not limited to the buttons the vendor gave you. You can clone the behavior of another entire platform.
12. Treat OpenClaw as infrastructure, not a toy
If you are installing OpenClaw, you are probably trying to build something real โ a platform, a tool, a personal assistant that matters. So treat it that way. Change the code. Run experiments. Get a second server if you need separation between production and play.
Ariel runs three servers plus a laptop install, each with its own job. That sounds like overkill until the first time a bad experiment does not touch the production box.
๐ฅ Roast Corner
The default permission mask exists because the average user will absolutely let the agent delete something important and then blame the tool. Unlocking it is correct for power users, but it is also the software equivalent of taking the safety off a gun. The fact that it feels better is not the same as it being safer.
The “three agents” rule is really a rule about human discipline. If you are the kind of person who cannot resist asking your one super-agent to do everything, you will eventually ask it to do something in the wrong context and it will obey beautifully. Agent separation is not architecture; it is a self-control hack.
And yes, running multiple OpenClaw instances on multiple servers is overkill for most people. Ariel gets away with it because this is literally his business. If you are doing it just to feel like a supervillain with a server farm, at least be honest about the hobby budget.
๐ค AI for Humans
OpenClaw is a power tool that ships with the guards on. Tip 2 is the part most people skip: they want the power but not the responsibility of choosing the safety setting. That is like buying a table saw and complaining that the blade guard slows you down. The guard is not the enemy; your impatience is.
The best practical advice here is the separation of agents. Most people think of AI as one super-assistant that does everything. A better mental model is a small company: a chatty front-desk person, a careful sysadmin, and a focused developer. You do not let the receptionist reconfigure the router. You do not ask the developer to book your flights. Same logic.
Memory is the part that makes agents feel human. Humans forget details; an agent with good notes does not. The unsettling part is that this memory is only as private as the machine it lives on. If your chat history becomes a searchable database, make sure it is your database.
Published 2026-07-14 from the YouTube video by Ariel Rubinstein.
๐ฌ Comments