It would be helpful if you could define “useful” in this context.

I’ve built a number of team-specific tools with LLM agents over the past year that save each of us tens of hours a month.

They don’t scale beyond me and my six coworkers, and were never designed to, but they solve challenges we’d previously worked through manually and allow us to focus on more important tasks.

The code may be non-optimal and won’t become the base of a new startup. I’m fine with that.

It’s also worth noting that your evidence list (increased CVEs, outages, degraded quality) is exclusively about what happens when LLMs are dropped into existing development workflows. That’s a real concern, but it’s a different conversation from whether LLMs create useful software.

My tools weren’t degraded versions of something an engineer would have built better. They’re net-new capability that was never going to get engineering resources in the first place. The counterfactual in my case isn’t “worse software”—it’s “no software.“

At my work ~90% of code is now LLM generated. It's not "new" software in the sense that you're describing but it's new features, bug fixes, and so on to the software that we all work on. (Although we are working on something that we can hopefully open source later this year that is close to 100% LLM generated, and I can say, as someone that has been reviewing most of the code, is quite high quality)

Well, on the surface it may seem like there’s nothing being created of value, but I can assure you every company from seed stage to unicorns are heavily using claude code, cursor, and the like to produce software. At this point, most software you touch has been modified and enhanced with the use of LLMs. The difference in pace of shipping with and without AI assistance is staggering.