Agents write units of changes that look good in isolation. They are consistent with themselves and your prompt. But respect for the whole, there is not.

Hacker News is a community-driven platform for sharing and discussing technology news, startups, and programming-related topics. Through user submissions and comments, Hacker News offers insights into emerging technology trends, industry developments, and entrepreneurial ventures. Readers can participate in discussions, share their insights, and stay informed about the latest advancements in technology and innovation.

Hacker News

After two years of using AI coding agents, the author discovered that while AI-generated code looks good in isolation and passes code reviews, it creates "slop" when viewed holistically. AI agents lack respect for overall codebase structure, architectural integrity, and neighboring patterns. Despite impressive individual pull requests, the cumulative result was unmaintainable code that couldn't be shipped to users. The author returned to manual coding, finding it faster and more productive when accounting for total cost, not just code generation speed.

After two years of vibecoding, I'm back to writing by hand

I think if you go to vibecoding and then back, you miss a lot. There is a whole universe of orchestrating agents and persistent knowledge that addresses exactly the shortcomings mentioned by the author. It’s really not clear at this point whether these techniques will stay as economically viable as they are today so now is the time to try it out and see what works. (By economically viable I mean the users who rely on low prices made possible by huge investments. At this moment the providers are anything but sustainable.)

Great post. “The AI had simply told me a good story.” Exactly right. Cal Newport recently said that the model, under the hood, <a href="https://www.youtube.com/watch?v=HB2Qx25Covo&amp;t=1878s" target="_blank" rel="noopener nofollow">“is trying to finish the story that you’ve given it”</a>.
The next 5 years are going to be quite interesting; I feel we’re creating a swath cognitively disabled people and we’re going to have deep regrets, much like we have now that social media has infected every element of our lives/society/politics/policy.
I’m continuing to sharpen my cognitive abilities. AI is a secondary or even a tertiary element to my workflow.
It seems there’s a fad to reverse this order, but there’s no free lunch! With social media, you are the product, and pay with your attention. With AI, you are the product, and you pay with your cognition.
I don’t vibe code, never really have and never really will. I delegate instead. Brain first, AI second.
<a href="https://cheewebdevelopment.com/dont-vibe-code-delegate-responsible-development-with-llms/" target="_blank" rel="noopener nofollow">https://cheewebdevelopment.com/dont-vibe-code-delegate-responsible-development-with-llms/</a>

I found that using AI in browser just to assist is very effective to me. I keep overal design in my head, I can navigate the code and shape it in a way I want but AI is helping me to do review, to generate some methods here and there or creates some repeating patterns to speed up. But I am trying to make these parts small enough so I can review them and understand. Using AI in a way that generates whole applications is not sustainable. It creates a lot of unnecessary complexity and code is usually a messy place where debugging or fixing bugs is very hard to do, especially when most of the code is AI generated and you don’t have the mental model in your head.

That’s the AI trap. It sells you believable fantasies that they can do amazing things but it always fell short at the end. If this trend continues, AI simply just gives more work for future good engineers to solve.

That’s absolutely right.
Vibecoding makes myself confusing in truth.