How to Set Up an AI Command Agent for Your Business
Most businesses meet AI one tool at a time. A chatbot on the website, a writing assistant for posts, an AI step inside a workflow. Each one works on its own, and none of them knows the business. Every new tool needs the same explanation again, and nobody can see how the pieces connect.
In the Lagoon Galaxy System™ I start from the other end: one command agent at the center, which I call the Core. It learns the business once, works across your computer and apps, and coordinates everything else. Here is how I set one up for a client.
1. Give the Core its own workspace
The Core only sees what you allow it to see. Before anything else, give it a dedicated computer, a separate user account, or at least one clearly scoped folder. Keep personal files, legal documents and anything unrelated to the work out of reach. This is the first and simplest layer of security, and it is the one most people skip.
The platform matters less than the setup. I build most Cores on Claude Code or the Claude desktop app, because they can work with files, folders, a browser and connected apps from one place.
2. Pack the flight kit
A flight kit is a small set of plain-text files in the Core’s root folder. They don’t make the agent smarter. They make it yours.
- IDENTITY — who the agent is, how it speaks, and what it never does (send, publish, pay or delete without your OK).
- CONTEXT — who it works for: the business, offers, pricing, clients, tools, goals and brand rules.
- LOGBOOK — one dated line for every correction or approval, so the same mistake never happens twice.
- Playbooks — one short recipe per repeatable task (“weekly newsletter”, “lead follow-up”), each ending as a draft.
The Core can interview you and fill most of these in itself. Your job is to correct what it gets wrong.
3. Give missions, not steps
A command agent is not a macro. Instead of listing clicks, give it an outcome and the boundaries: “I want newsletter open rates up. Review the last ten sends, find what the top performers have in common, and tell me what you’d change and why. Don’t send anything.”
The formula I use for every mission: the goal + what it may use + what it may not do without me + leave it for review.
4. Connect apps one gate at a time
Connectors let the Core enter your email, calendar, CRM or files and work on real data. Open only the connections a task needs, and set what each one can do. Reading and drafting are fine on day one. Sending, publishing, moving appointments and anything involving money waits for a person. I cover this in detail in the AI permissions guide.
5. Review the signal every day
The Core gets better through feedback, not through more instructions. Before you accept a result, ask three questions: what did you assume, what couldn’t you verify, and which part is weakest? When something is off, correct it with the reason and have the Core log it in the logbook or the playbook.
Research on AI-assisted knowledge work by Harvard Business School and BCG (2023) found large gains in speed and quality on tasks AI handles well — and clearly worse results when people accepted AI output without checking it on tasks it handles badly. The review loop is not optional. It is the system.
6. Add one mission at a time
Ten missions on day one means none get reviewed properly. Start with the single task that repeats the most and ends as a draft. When it comes out right consistently, add the next one. Here is how to pick that first task.
What you end up with
After a few weeks you have one agent that knows your business, a set of playbooks that capture how your team works, and a clear record of every correction. From there, the Core can launch project agents for specific builds and help run the automations that must work the same way every time.
If you want this built for your business, CTRL+R Studio™ sets up the Core, the flight kit and the first missions as part of our AI Systems service.