80% Done 80% Well

No doubt you’ve seen that blog post making the rounds this week. I read it this morning, or I should say I “read” it, in that I read down to the the action items, then sort of skimmed it from there. I didn’t need to hear someone say I should learn to use coding agents. I already use coding agents, usually when I need them to do something tedious.

But the original poster’s breathless description about how effective these new coding agents released a week ago are made me curious. I don’t think I’ve tried having a coding agent build something from scratch, completely unsupervised. So I fired up Claude Code and asked it to build me two things: an RSS reader, and a replacement for My Tracks, both in Flutter targeting Android.

I’ll spare you the details (although I may write more about it in a later post) but my impression on the capabilities of these models is… well, I will say that it’s ability to build an entire project from scratch is pretty impressive. What’s less impressive is that I found it needed further intervention from me to get it working and up to spec. The RSS reader didn’t load the feeds I tried it, and once I fixed that, it got itself into an infinite loop. The My Tracks replacement was better, and I think it works (I’m always skeptical of location tracking without excessive testing), but there were some questionable visual decisions that require fixing.

So after using these models to get 80% of what I wanted 80% done well, I tend to side towards the takes made by others writers regarding this post. It’d be foolish for me to say that AI won’t change things: I feel pretty confident that that’s going to happen. But going to a world where unsupervised AI agents will be making apps all the apps for all of us? There I remain skeptical.