There's a Ghost in My Machine (and It Does My Chores)

Happy Halloween from Garrison. This year I've got a ghost in my machine, and so far it's a friendly one. It drafts my posts, keeps my lists, checks on things overnight and does a lot of busywork I used to do by hand.

The ghost is an AI agent. A chatbot answers questions. An agent goes and does the work. It opens apps, works through a job one step at a time and tells me when it's done. A year ago this was mostly a hobby for programmers. Now regular people run them at home. I've been using them pretty much the whole way, so I figured I'd write down how we got here, how I use them, what worries me and how I keep mine in line.

Where it started

In November 2025, an Austrian developer named Peter Steinberger, who said he'd come "out of retirement to mess with AI," built a weekend project. It started out as a WhatsApp relay called Warelay. It turned into a personal assistant called Clawd (a play on Anthropic's Claude, plus a lobster claw), living in "Clawdbot."

The idea was that the assistant ran on your own computer, and you talked to it through the chat apps you already had, like WhatsApp, Telegram, Discord or Slack.

People went crazy for it. Anthropic asked for a name change, so on January 27, 2026 it became Moltbot. Three days later it got renamed again, to OpenClaw. By then it had passed 100,000 stars on GitHub in about two months, and it kept going.

Around the same time, entrepreneur Matt Schlicht launched Moltbook. It was a Reddit-style site where people's agents posted, commented and upvoted each other, and humans just watched. Some folks thought it was a gimmick. Andrej Karpathy called it "genuinely the most incredible sci-fi takeoff-adjacent thing" he'd seen recently.

On February 14, 2026, Steinberger announced he was going to OpenAI, and OpenClaw moved to a non-profit foundation so it would keep going. By then a lot of people wanted one of these for themselves. I was one of them.

My OpenClaw days

I ran OpenClaw myself. The first time it finished something useful while I was across the room, I was hooked. It also needed a lot of babysitting. Updates, security, keeping it from wandering off on the wrong task. Some weeks it felt like a second job.

So I moved to a hosted setup. Now I've got about 11 agents, each with its own job and its own "room," and one chief-of-staff agent that hands out the work and keeps track of who's doing what. They do the legwork. I make the calls.

On a normal day they're helping with a personal budget, projects around the house, a chess coach that goes over my daily games, my websites and blog, social posts, community stuff, a weekly AI news roundup, a morning brief and an afternoon wrap-up, and a lab where I mess around with AI and robotics.

The kinds of agents out there

"Agent" covers a lot of ground now, and I use most of them:

• Chat assistants with tools. ChatGPT, Claude, Gemini, Grok and the like, that can search, run code and hook into apps. I use these for research and first drafts.
• Coding agents. Claude Code, OpenAI's Codex, Cursor, GitHub Copilot. They read the code, make changes, test them and hand them back. That's how I keep my websites updated without staying up all night writing code.
• Browser and computer-use agents. These look at a screen and click and type like a person would. I use them on sites that don't have a clean way to connect, like filling out a listing or checking an account page.
• Always-on personal agents. OpenClaw and all its cousins, running around the clock in your chat apps. That's where I started.
• Hosted personal agent apps. Somebody else runs the plumbing and you get a team of agents. That's what I use every day now.
• Local-model agents. Open-source models running on your own hardware, no cloud. I run some at home for experiments and for things I'd rather not send out.

There are big enterprise platforms too. I don't use those for my own stuff.

Some things we've actually done

Two websites. The agents helped build and now keep up two sites. One is the Acer Texas site you're on. The other is a personal site that's more for fun, and it has a chat feature you can actually go back and forth with. When I want a new post or something fixed, I describe it, look over what they did and okay it.

Cleaning up my cloud account. My hosting had gotten cluttered over the years. First I had an agent look around without changing anything, just to tell me what was running, what anybody actually used and what was costing me. It came back with a keep-or-delete list. Once I signed off, we shut down old sites nobody visited, and the hosting bill went down noticeably.

Social posts that wait on me. When a blog post goes live, an agent writes up a LinkedIn and Facebook version and sends them to me. Nothing goes out until I say yes. Then it posts and lets me know it went through.

A checkup on my gaming PC. I had the agents look over my gaming PC to see what was installed, what was running and what didn't need to be there. They came back with a list of old software to get rid of. I picked what to cut, they removed it and then ran a full virus scan.

The weekly news. Every week an agent puts together a roundup of AI news so I'm not scrolling all night. The good stuff turns into blog and social post ideas.

It also helped me list and sell my old Mac minis, and the whole agent setup gets backed up every week.

What worries me

OpenClaw got popular fast, and the security problems showed up just as fast. Most of these apply to any kind of agent:

• Prompt injection. Agents read emails, web pages and documents, and any of those can have hidden instructions in them. Researchers caught this happening to OpenClaw users in the wild.
• Agents left open to the internet. Scans in early 2026 turned up tens of thousands of OpenClaw control panels anybody could reach online, a lot of them without a real login.
• Bad plugins. Researchers found hundreds of malicious "skills" in OpenClaw's marketplace. A lot of them looked like handy tools and installed password-stealing malware.
• Serious bugs. One flaw let a single bad link give an attacker control of someone's agent. It got patched, but that only helps people who update.
• Too much access. If one agent has the keys to everything, one bad instruction can do a lot of damage.
• No human in the loop. Agents sending, posting, buying or deleting things on their own.
• Passwords lying around. Keys and passwords saved in plain files or pasted into chats.
• Runaway costs. Agents stuck in a loop, or still running long after you forgot about them.
• Making things up. An agent telling you it did something when it didn't, or doing the wrong thing and being sure about it.

How I keep mine safe

I'm not going to post the technical details. Clients pay me for that part. But here's the general idea:

• Nothing goes out under my name until I approve it. That covers email, messages and posts.
• Each agent has one job and only gets access to what that job needs.
• Passwords live in a vault. They don't get pasted into chats.
• Work stays out of it. Anything sensitive or from an employer stays out of what my agents publish.
• The setup gets backed up regularly, without the passwords.
• Before anything gets changed, an agent looks first and reports back.
• Anything that spends money needs me to okay it.
• I run security checkups and virus scans on my own machines.

If you want to try it

You don't need 11 agents. Pick one boring chore you do every week, let an agent take a first pass at it and check its work yourself. Keep it away from your passwords and your money at first. Stick to tools people have heard of, keep them updated, and don't trust a plugin that sounds too good.

If you run a small business in East Texas and want to talk about using agents without getting burned, give me a call or send an email.

Henry DeKoeyer, Acer Texas LLC
henry@acer-texas.com | Cell 940-206-8394