Key takeaways
- Keep it simple: one server you own, reached from your desk and your phone through Claude or Codex.
- Every agent drafts and a person approves; new automations ship switched off.
- One registry file is the only source of truth for where every app lives.
- Edits go through a queue so two agents never change the same files at once.
I run an agency with a small team and a lot of AI agents. If you own a business, you've probably seen the posts that make this sound like a machine that runs itself while the founder sits on a beach.
This is the version without the beach. What the agents actually do, where the lines are, and what broke along the way.
One server, two doors
The first thing that made this work was keeping it simple: one server I own, with everything on it. The software the business runs on, the data, the memory, the skills.
- From my desk. Claude or Codex on my laptop, working on that server. They build and fix the software the business runs on: Ad Blitz OS, this website, the automations behind both.
- From my phone. The same setup, the same memory, the same rules. I can check a number, approve a draft or fix something from anywhere.
I tried a more complicated setup: a separate team of agents on another platform, each with its own job. It worked, but it was one more thing to keep in sync. One server with two doors is easier to trust, because every agent reads the same facts.
| Who | Where | What |
|---|---|---|
| AI (Claude or Codex) | On my server, from desktop or phone | Builds, fixes, drafts, watches |
| Watchdogs | On the server, every few minutes | Check the apps, disks and backups |
| Me | Wherever I am | Decisions, approvals, client calls |
Nothing sends without a click
Every agent drafts. I approve. That rule covers emails, texts to clients, ad changes and anything that moves money.
It sounds slow. It isn't, because the drafting is the slow part and the agents do it. My part is reading and clicking. New automations ship switched off, and I turn them on once I've watched them do the job right on real work.
The one place this bends is lead follow-up. Katie, the AI receptionist in Ad Blitz OS, replies to a new lead in seconds, because a lead that waits an hour is mostly gone. Even there, she follows scripts I approved, and the conversation lands in one inbox my team can see.
What broke
An agent that trusted an old list. For months an earlier agent, Hermes, sent me "app down" alerts that weren't true. It was reading a list of where each app lived, and the list went stale when we split the servers. I retired it on August 31 and made one registry file the only source of truth for every app. Every agent and every watchdog reads that file now.
Two agents in the same file. With several agents building at once, two of them would edit the same app and undo each other's work. Now every edit goes through a queue: an agent claims the files it will change, the others wait their turn, and nobody pushes over someone else's half-finished change.
Writing that sounded like an agency. Early drafts had the tone I hate: hype words, long dashes, promises. The fix was memory, not a better prompt. Every correction I make becomes a rule the agents read before they write, and after a few weeks the drafts sound like me.
The quiet part
Most of what the agents do, I never see. Watchdogs check the apps every five minutes, the disks twice a day and the backups every morning. When everything is healthy they say nothing. When something is wrong, one message reaches my phone.
That's the goal of the whole setup: the business runs, and my attention goes to the few things only I can decide.
If you want this for your business
Start smaller than you think:
- One system on one expensive job, usually lead follow-up. My first five AI skills are a good place to start.
- Drafts only until you trust it.
- One place for the truth, so every agent reads the same facts.
The setup I use is in the AI OS Starter Kit, and if you'd rather build it with me, book a consulting call.
Questions people ask
Do AI agents run the agency on their own?
No. They draft, build and watch. I approve anything that sends, spends or changes a client account, and I make the decisions.
Do I need a team of specialised agents?
No. One server with Claude or Codex, reachable from your desk and your phone, does the job. A separate agent platform is one more thing to keep in sync.
Where should a business owner start with AI agents?
One system on one expensive job, usually lead follow-up, with drafts only until you trust it.
Sources
New writing, straight to your inbox.
Software, AI and ad systems, from the build. One email when something new is up. Unsubscribe any time.