Claude code’s harness isn’t the best, but it’s still better than just promoting the model directly. One thing cursor has going for it is the vector embeddings of your codebase to make the harness more token efficient, but I still switched from cursor to managing parallel Claudes in cmux a few months ago and have had little reason to go back.
Have you tried composer-2.5? Supposedly it’s fast, decent, and far cheaper than sonnet.
Same, but only for ~15 years. I’m amazed at how most of this thread is people saying it’s useless.
Agentic coding (Cursor, Claude code, etc) is an absolute game changer, especially when paired with strong models like opus 4.6.
I used it yesterday to implement a feature I knew exactly how to implement by hand, but it did perfectly in 10 minutes what would’ve taken me at least an hour or two if I had to implement it myself.
My one fear is skill atrophy, I still take certain tasks manually when I have bandwidth just to keep my skills sharp. That’s really the only reason though, modern models are capable of pretty much anything I would throw at them.
I used to have a near maxed out 2019 i9 mbp. The 2020 base model m1 blew it out of the water performance-wise.
Granted, a big part of that was the apple silicon, but the i9 was supposed to be a powerhouse. It just wound up spending most of its time thermally throttled.