Tools, workflows, dead ends. No hype, no forecasts; just what I've built, used and in part thrown away again myself.
I build almost all my projects with AI agents now, and since I started doing that, they've become a great many more.
It's not just autocomplete: I hand over a task and judge what comes back. Sometimes by reading it, sometimes by trying it out, increasingly through checks I build for the job first. At some point a model will probably take over the judging itself.
The hard part isn't the building, it's the keeping. What I learn about an agent this week is outdated or forgotten in three months, and most of it is written down nowhere, because it's too new. This blog is the countermeasure: write it down before it rots.
The examples come from projects I built or was responsible for myself: from my work, from my private life and everything in between.
My son can't get out of Cyberpunk 2077, because there's always one more quest getting in the way. I couldn't get out of a small feature that rebuilt half my server over one morning. Modern games have done away with the save point, and AI does the same when you code with it.
For months I've had AI agents taking logs and markets apart. At some point I turned the same machine on myself, on my heart rate, my money, my own decisions. It doesn't just show me data, it grades me. And the uncomfortable part is how often it's right.
My wife says the midlife crisis is knocking at my door, because I bought a gravel bike. The research for it ran through an AI agent first. Looking back, it was more on point than I was, and in the end I bought from the shop it could see the least.