Why software keeps getting worse even though AI writes the code now

AI writes code faster than any of us thought possible a couple years ago. Autocomplete turned into full agents that ship whole features while you grab a coffee. So here is the part nobody wants to say out loud: the software on your phone, in your car, and at your bank keeps getting worse anyway.

I read a post this week called "Nothing Works and Everyone Is Euphoric" that put words to it. It hit the Hacker News front page and pulled 450 points and 369 comments. That is not a hot take dying in the void. That is a room full of engineers nodding along, saying yeah, my banking app makes me FaceID three times, my car's screen reboots mid-drive, the warranty form fails at the very last field.

If coding is basically solved, why is all of this getting worse?

Faster code, same broken product

Here is my honest take. Writing code was never the real bottleneck for good software. It was the cheap part that only felt expensive. AI made that cheap part nearly free, and everybody celebrated like the hard part got solved too.

It didn't. The hard part is deciding what to build, saying no to the ten features nobody asked for, and actually fixing the boring bug that makes people hate your app. None of that got faster. If anything, cheap code made it worse, because now you can ship the wrong thing at ten times the speed.

Nobody gets promoted for fixing bugs

Software teams chase KPIs, and a stability fix does not demo well. No product manager stands up in the quarterly review and says we shipped zero features and redesigned nothing, we just made it stop crashing. So that work quietly does not happen, and the rot compounds release after release.

Bloated SaaS is where this hits hardest. These products have to keep looking alive to justify the subscription, so they bolt on features forever. Every new tab, every AI sidebar you never asked for, is more surface area to break. You end up paying monthly to be the unpaid tester for stuff you never wanted.

More of the wrong tool, just faster

That is the whole trap. AI did not make companies build the right tool. It made them build more of the wrong one, faster. The bank app, the fridge warranty site, the car OS, none of those are bad because someone couldn't write the code. They are bad because the thing was overbuilt, over-scoped, and nobody owned the job of keeping it simple.

Speed only helps if you are pointed at the right target. Point a faster tool at the wrong target and all you do is arrive at "worse" sooner.

The case for lean, custom software

Here is why I am actually kind of excited. The same tools that let a big company pile on bloat also let one person build the exact right tool for one business. Small and tight beats big and bloated, and it always did. AI just made small and tight cheap enough to be worth doing.

When I build custom software for a business, the whole point is that it does the handful of things they actually need and nothing else. No 200 settings screens. No feature they will never touch that still finds a way to crash. Less surface area means less to break, which is pretty much the actual definition of software that works.

That is the opposite of the SaaS model. Off-the-shelf tools are built for the average of ten thousand companies, so they fit none of them well and they carry the weight of all of them. A right-sized tool built for you is smaller, calmer, and it keeps working after the update lands.

What I would build instead

If your software keeps getting worse, the fix usually is not a better subscription. It is less software, built on purpose. Map the five things your team does every single day, build a clean tool that does exactly those, own it, and stop renting bloat that gets shakier with every release.

Coding got solved. Judgment did not. That gap is the whole reason lean, custom software is about to have a really good decade.