Honestly agents can't one shot shit except cool demos. I mean in a way they can but I let an Astra/Fable 5.1 duo build an android app and figured I'd really just loosely vibe code this for once instead of supervising.
It's a working pile of trash that has quadratic growth in the remote postgres database resulting in 10GB stored in a few days for what should have been like 5 MB of data.
It's not a skill issue, I know how to get good results but I figured maybe I don't need to be as cautious as I am all the time.
Not that I don't believe we can get there eventually.