When I was a kid, my grandparents kept a bound collection of Nauka i Zhizn (Science and Life), a Soviet popular science magazine. At some point, I read a short piece about a programmer who wrote programs that modified themselves. The idea went in and didn’t come out.
I spent the next thirty years mostly doing other things.
Then AI agents became real, and I’ve spent the last few months building what is essentially a homelab that thinks. Three refurbished office PCs in a Proxmox cluster — a hundred and fifty euros each, give or take, picked from a dealer who recycles corporate fleet hardware. Forty-odd Linux containers, one per service. Tailscale threading everything into a network that doesn’t touch the public internet. Hermes as the agent harness. Hindsight storing what the agent accumulates about me across sessions. Langfuse watching every token.
I talk to it through Discord. It remembers almost everything.
In May, I sat on a video call for 94 minutes and explained all of this to six people. The format was informal — part show-and-tell, part interrogation. We had transcriber running.
I was explaining the stack. The memory architecture. How I chose one service over another. A script that monitors US Senate financial disclosures. An Oura data pipeline that does in thirty seconds what the Oura app doesn’t do at all. The diagnostic I use to tell whether an agent is working or flailing.
Near the end, someone asked me to send them two things: the senate trades setup, and a skill that generates a fortnightly digest of forgotten brainstorms. I said: I didn’t write those down — I’ll check the transcript later.
The transcript was in my workspace right after. The agent read it. By the time I looked, there was a task in my project tracker: Send Yuri: senator trades + braindump digest skill.
The task was created by the thing the meeting was about.
The infrastructure isn’t sophisticated. There’s nothing here that isn’t open source, nothing that costs more than three cheap computers and a Claude subscription — a subscription that Anthropic will probably restrict in a week or two, at which point I lose maybe twenty percent of the functionality and rebuild the rest in a few weekends. I’ve thought through what breaks and what doesn’t. The dependency is shallower than it looks.
The thing that makes it work isn’t the stack. It’s the context the stack accumulates. After enough use, it knows where things live. It knows what matters. You say go, and it goes. You don’t write instructions — you’ve been writing them all this time without realizing it, every time you said no, that tool exists, use this one, every time you corrected it and didn’t have to correct it again.
The diagnostic is embarrassingly simple: watch the command log. When an agent is confident, you get four lines. List skills. Select skill. Run. Return. When it’s lost, you get twenty — wrong guesses, SSH attempts, trying to find a workaround because it doesn’t know the right path. The number of lines is honest. Agents don’t hide their confusion, they thrash.
Someone asked how I got into this. I said: childhood, Soviet science magazine, programs that modify themselves. The idea stuck. I’m happier building this than almost anything I’ve done professionally, which probably says something, though I’m not sure what.
Someone else asked if I was trading on the senate signals. I’m not, I’m passive, I look, I use it to inform decisions occasionally. But I built the thing in an afternoon. The idea came first and the infrastructure was already there.
That’s the shift: the gap between I want to know when senators are buying and selling and I know when senators are buying and selling is no longer skill or time. It’s just whether you asked. You stop thinking about whether you can build it and start thinking about whether it’s worth building. Different problem.
The transcript from the meeting is 9,400 words. This article is about 900. The agent turned the call into a task, and I turned the task into the article, and the article describes the system that did both.
The programs that modify themselves, it turns out, mostly modify you.