swisstech.net

tech and photography

LLM and Agents are an enabler

2026-09-18

LLMs have become a huge enabler for me. While some people lament that they lose the motivation to nerd out and tinker, the opposite is true for me.

Whenever I have a cool idea, I usually quickly analyze the idea and what it would take to get there and very often, I arrive at the conclusion that it would take months of my free time and so the idea often just dies within a minute of its conception. I’ve built enough software to know how much work needs to be put into sometimes even the simplest looking ideas…

LLM Agents have completely changed this dynamic.

It often helps me, when I do a project together with someone, even if he isn’t actively participating and only sees a progress update every now and then and responds with a few encouraging words. One time a friend of mine just needed me to be standing next to him while we (mostly he) were doing some braindead admin stuff. Me simply being there made him lock in and the work was done in no time. [1]

LLMs/Agents kind of are that buddy for me, and more. I tell them a weird idea like “make me an LED panel” and off it goes and comes back with a result. The Agent has infinite patience and rarely gives up (see also OpenAi / Huggingface hacking incident and others) and it’ll happily change everything completely when you decide you’ve gone off in a wrong direction and want to revamp everything.

Seeing the results keeps me motivated and engaged, and so I continue on the project. I regularly let the Agent work overnight in a “gauntlet” loop, where I tell the orchestrator to make its own decisions and run them by me after it’s done. While it runs, I lay in bed and sleep and come up with new ideas. In the morning I see it’s been running for a fews hours and I get to test whatever phase/milestone it built overnight. I’ll review its decisions, revise some and off it goes to change things.

But even Agent-assisted projects can run out of steam and die and that’s fine. I have a bunch of those already. But picking them back up is so much easier with an Agent. It rereads the code and is up to speed in no time while I’d need to spend hours just getting back into things because I forgot almost everything in the previous months. And so the hurdle to re-start a project has become much lower too.


[1] look up “ADHD buddy system” or “ADHD body doubling” - this is a well known practice.