TypeScript.fm - The Friendly Show for TypeScript Developers · Hands-On Engineering Podcasts · November 2026
Erik introduces a conference talk by an engineer who let AI agents refactor a codebase for three weeks and then read the result. He says it matches his own experience that agent-driven code turns into a mess.
right. Big friendly giant. Big friendly deal.
Love it.
So there's a really interesting video that I watched that I do agree with. And it kind of aligns with the whole premise that I put in No Sloptober, which is the, I'm trying to coin my own software engineering law, Onerheim's law, that agents can only increase entropy. Humans are the only things capable of decreasing it. I'm not saying that humans always decrease entropy. But I think that's exactly what happened to this fellow in this video by AI Engineer. It's by KG Kanazawa at Microsoft. And they talk about how they went with the vibes and the vibes didn't go so well. And it became a whole giant mess.
Yeah. And to be clear, if your reaction to that is, well, I can prompt the agent to reduce entropy. That's exactly what Eric means.
You need a human
in there because the issue is that if you let the agent left to its own devices, left on its own Ralph loop or whatever you're doing, it will just increase entropy. It will never remove entropy.