Isn’t substack a nazi bar

Worth calling out 3 years ago SWEs were also 10-15 years out from being automated.

Things change very fast nowadays.

Medicine has no “undo” button. As with self-driving cars, the early promises were probably functional enough, buuut we’re not gonna bet human lives on it. Even the scum-sucking billionaires are so terrified of liability that they’ll insist ‘the human driver was given three hundred milliseconds of heads-up, so they drove the car into that wall!’ Frankly, a model could detect 99% of tumors with a 1% false positive rate, and we’d still hyper-fixate on that gap. Humans are narrative animals and statistics are not real to us.

But nobody gives a fuck about bad code. There’s places it matters - yet those places are already getting hacked left and right, week after week, thanks to bad code written entirely by authentic meatbags. Mere programs on your home computer having a fucksy-wucksy is practically a non-event. It used to be even less! In the 90s you just expected program to crash when you did certain things and learned not to do those things. As late as Skyrim people were shrugging off the possibility of losing a save file by suggesting you rotate through several. The stakes are low and the workarounds are largely painless.

Perhaps more importantly - code is rich output. Automated radiology is mostly ‘cancer y/n’ or drawing a red circle and saying ‘that ain’t right.’ Automated driving is three one-dimensional analog controls which need need updates constantly and immediately. Code can be anything. Code can go from a one-sentence todo of intent to a sprawling mess of interdependent files. So improvements don’t mean narrowing the number of missed tumors or hit children - new things become feasible.

Recently frontier models went from ‘probably won’t fumble a shell script’ to ‘refactored three commercial binaries into one multiplayer browser game.’ The Fuck_AI crowd (the mods, no less) have sneered ‘that’s not reverse engineering!,’ which is at best insisting that submarines cannot swim. Someone still wanted a hideously complex thing accomplished, and the robot did it for them, in like, a day.

That’s coming to small models. We will take it for granted.

You can’t seriously believe it ends exactly here.

agentic coding has been very good at easy things for a while now … Even Anthropic reported late 2025 that much of their productivity gain comes from non-critical, nice-to-have code.

Disruptive innovation always starts from the bottom. It’s a time-saver for low-quality busywork, freeing up resources for profitable and satisfying challenges. Then it’s a good-enough median option, leaving the artisans with high margins for high quality. Then the old ways can’t compete anymore. We are at step two.

Software engineers’ productivity does not matter - they’re facing irrelevance. The communication bottleneck between engineers and customers only exists because the customers’ goals require the engineers’ expertise. That goes away if the server itself can turn tickets into commits. People who want custom programs can potentially get it by asking their own computer. It’s software as a service, as software.

Nobody is suggesting agentic coding will result in less software. But the people making it won’t be software engineers as this author defines them. The type of people who treat Excel like a database are going to roll their own everything. Which is, without sarcasm, the long-awaited realization of everything home computing was invented for. God help us all.