YEP

Things are changing. Now everyone can build bespoke apps. Are these apps pushing the limits of technology? No! But they work for the very narrow and specific domain they where designed. And yes they do not scale and have as much bugs as your personal shell scripts. But they work.

But let's not compare these with something more advance - at least not yet. Maybe by end of this year?

We switched from Sonnet 4.5 to Opus 4.5 as our default coding agent recently and we pay the price for the switch (3x the cost) but as the OP said, it is quite frankly amazing. It does a pretty good job, especially, especially when your code and project is structured in a such a way that it helps the agent perform well. Anthropic released an entire video on the subject recently which aligns with my own observations as well.

Where it fails hard is in the more subtle areas of the code, like good design, best practices, good taste, dry, etc. We often need to prompt it to refactor things as the quick solution it decided to do is not in our best interest for the long run. It often ends in deep investigations about things which are trivially obvious. It is overfitted to use unix tools in their pure form as it fail to remember (even with prompting) that it should run `pnpm test:unit` instead `npx jest` - it gets it wrong every time.

But when it works - it is wonderful.

I think we are at the point where we are close to self-improving software and I don't mean this lightly.

It turns out the unix philosophy runs deep. We are right now working on ways to give our agents more shells and we are frankly a few iterations there. I am not sure what to expect after this but I think whatever it is, it will be interesting to see.