By Rolf Versluis · Published September 1, 2026 on Uncommon Knowledge in Business and Life

It's 2am when my phone buzzes. I don't even get out of bed because I know whether I care before I look at the face in the picture from the farm camera. Telegram notification. Someone's on the property. The agent wrote that program, and I texted it a request a few months ago and we went back and forth until it worked.

That's what I want to explain, because the word "agent" is doing a lot of work right now and most of what it gets used for isn't what I mean.

I talk to my agents through Telegram, and that's the only interface I need.

Each one has a profile that explains who I am and what they're for, and they run while I sleep and pick up where they left off. I don't text them every morning, which means they keep working on their own.

I put a Hermes agent on a Mac Studio. It runs a large language model, the same kind that powers ChatGPT. I don't open a chat window to ask it things. I write a job description into its profile and it figures out the rest. When it reads the job it decides what to do, uses tools to do it, then comes back with the result.

ChatGPT and Grok are useful when I'm at my desk. Close the tab and they forget everything. I run mine locally, and that's the difference. My data stays on my machine. Health care and financial services, anything with sensitive data. That's why I bother running it on my own hardware.

I wrote the profile myself, and mine says I'm an engineer in Alpharetta, and it lists my businesses, and says I'm working on a chicken farm in Chatsworth and a small AI consulting practice. Also says I'm working on a wooden sword audiobook and a YouTube series and about six other projects I haven't finished yet.

I give each one a job, and some do research while one helps me write code and another runs the surveillance system on the farm.

For new projects I don't open a chat window. I text my agent directly, and sometimes I record a voice memo while driving, and sometimes I type a few sentences in Telegram. It reads what I sent, decides what to do, and either starts working or asks a clarifying question.

That's how it works.

I bought an IP security camera for the farm and mounted it on the building with a simple ask. When somebody drives onto the property, I want a notification with a picture so I can tell if I know the person.

We went back and forth for a few days while the agent looked for an open-source project first. There's a leading one called Frigate, and it would have worked mostly, but it didn't do what I wanted, so the agent built a custom one instead. It took a few months, and now when somebody drives onto my property, the system identifies them, takes a picture, and sends it to me on Telegram with no false positives. A branch blowing in the wind doesn't trigger it. A shadow doesn't trigger it.

I've used cell-phone-connected trail cameras before, and they send a picture every time anything moves, which makes them useless. Too much noise, not enough signal.

I couldn't hire a person to watch my security cameras 24 hours a day, but an agent can do it, and the program runs without pay, without sleep, without quitting.

The agent builds up skills, and a skill is a note it writes to itself about how to do something. The first time I had the agent install nginx, I walked it through every step. Now it's got a skill for installing nginx, and the next time I ask, it's 99% of the way there before it starts.

When I do something that needs repeating, I write a script instead of having the agent do it again. The agent is fast and the script is consistent. Different things matter.

After eight months, I've got hundreds of these skills. The work speeds up, the agent stops asking the same questions, and I stop explaining the same context. That's what accumulates.

The agent doesn't make decisions for me, and it won't spend money without asking. It can't answer the phone.

It takes the loops. The report I keep meaning to build, the follow-up email I keep forgetting, the vendor that needs chasing. Those are the loops the agent takes. All of it.

That's the work.

When I run into a problem I text the agent about it, asking how other people solve similar things, and it goes out and does research. Usually it comes back with a configuration fix or a tool I didn't know about, and more often the suggestion is that I'm going about this wrong and should try it this way instead.

For the detail work, the stuff I avoid, the agent handles it. I set up a script when the loop is big enough. I bring the judgment.

I run a small consulting practice with my son Grant, a software developer. We implement Hermes agents for owner-operators of small and medium businesses, covering North Georgia and the I-75 corridor. Alpharetta, Chatsworth, Chattanooga, anywhere in between.

If that sounds like something you need, send me a note. Thirty minutes, no pitch. If a Hermes agent isn't the right fit, I'll tell you that too.