I don't think he is saying agents are not useful at all, just that they are not anywhere near the capability of human software developers. Karpathy later says he used agents to write the Rust translation of algorithms he wrote in Python. He also explicitly says that agents can be useful for writing boilerplate or for code that can be very commonly found online. So I don't think he is saying they are not useful at all. Instead, he is just holding agents to a higher standard of working on a novel new codebase, and saying they don't pass that bar.

Tbh I think people underestimate how much software development work is just writing boilerplate or common patterns though. A very large percentage of the web development work I do is just writing CRUD boilerplate, and agents are great at it. I also find them invaluable for searching through large codebases, and for basic code review, but I see these use-cases discussed less even though they're a big part of what I find useful from agents.

I’m not saying he’s saying agents aren’t useful at all. It’s literally in the quotes I provided that he says they are useful for some subset of tasks.

I’m saying that he is answering the question “are agents useful at all”. not “can agents replace humans”.

His answer is mostly not. He generally prefers autocomplete. But they are useful for some limited tasks.

My biggest takeaway is that agents/LLMs in general are super helpful when paired together with a human who knows the inside and out of software development, who uses it side-by-side with their normal work.

They start being less useful when you start treating them as "I can send them ill-specified stuff, ignore them for 10 minutes and merge their results", as things spiral out of control. Basically "vibe-coding" as a useful concept doesn't work for projects you need to iterate on, only for things you feel OK with throwing away eventually.

Augmenting the human intellect with LLMs? Usually a increase in productivity. Replacing human coworkers with LLMs? Good luck, have fun.