Because everyone, including this forum, is addicted to the instant gratification of LLMs. It’s pure hubris of thinking you can scan the output and it does what you think it does.
TBH I don't really feel the same most of the time. I give the LLM little chunks to do. I read the code. I think. I plan. I write a bit of code. I have the LLM crunch out some bullshit task like setting up an annoying C repo. There aren't that many moments in building with LLMs where things line up so the AI can just absolutely nail some code and save me a ton of time.
I think a lot of people have a sort of “slot machine” experience with it at some point. You just start firing off prompts on some new project, wait a few seconds, see what prize you got. Then you start doing that over and over just letting the LLM code and code and not even review what it’s doing. It really is like getting hooked on gambling. You’re getting a thrill from anticipation, not the actual results.
This is what I personally consider “vibe coding”, not simply using LLMs or agents or whatever in your workflow
I’m confused by the people that “don’t even look at the code anymore”. I’m looking at the code Opus generates. And I’m glad that I am, because it needs revision.
Absolutely, vibecoding is addictive, agents being able to do all of these hard things that would otherwise take days or weeks to figure out how to do by hand.
It's just another avenue of dopamine addiction, not unlike scrolling TikTok, Reddit, or wherever except vibecoding is disguised as being productive.
> There aren't that many moments in building with LLMs where things line up so the AI can just absolutely nail some code and save me a ton of time.
Your workflow is similar to mine, and not "agentic". The proponents of Agentic workflows would have you handover the bulk of the edits to AI, and you'd hand-hold/course-corrrect its approach via chat while it does the thinking.
I tried the agentic approach for code-gen once, and found it mentally draining. Its like pairing with an over-enthusiastic junior on cocaine that can also type 2000 wpm.
Chunking small changes is great because you don't need the latest and greatest models, Flash variants are more than enough.