It sounds like my experience has been similar to yours. I have found a few places where agentic coding produces pretty good results.. generally very small patches that could have been written by anyone. I give the tool credit for finding small bugs that nobody noticed before.
On the larger or novel tasks I've thrown at these models, including some of the top tier options, the tools have either produced incorrect solutions, solutions written in a very inefficient way, or solutions that actually introduced more problems. I've taken some of these same challenges to other AI experts as who couldn't believe the tool failed. None of them were able to get good results either.
Everybody is desperate to carve out their slice of the AI Gold Rush right now before it all condenses down and developers realize they can't give up all agency to coding tools trained on the great mass of garbage that's out there. If at some point these tools truly do make developers 10x more efficient, they'll naturally get adopted. Hype chasers and product marketers are not the ones to listen to right now.