The Delegation Is the Hard Part

Andrej Karpathy, on how programming has changed in the last two months:

You're not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks in English and managing and reviewing their work in parallel.

His example: he wrote a single natural-language paragraph describing a local video analysis pipeline — inference server, web dashboard, system services, the lot — and an agent went off for thirty minutes, hit problems, solved them, and came back with a working system. Weekend project, done over lunch.

I believe him. The current generation of coding agents genuinely is a step change from six months ago. Anyone who's used them seriously knows this.

But look at the paragraph he wrote:

"Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me."

That's not English. That's expert shorthand that happens to use English words. Every clause compresses years of systems engineering knowledge into a delegation instruction that only works because the person writing it already knows exactly what "done" looks like. "Set up vLLM" is four words carrying about forty hours of context.

Karpathy, to his credit, buries none of this:

It's not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality).

That caveat paragraph is doing more honest work than the rest of the post combined, and it's the part nobody will screenshot.

Here's what I keep coming back to: the thing that makes agents useful — the ability to specify a task clearly enough that a machine can execute it — is the same thing that made senior engineers valuable before agents existed. The skill was never typing. It was knowing what to type. Agents don't change that. They just make the typing part fast.

This is a genuine productivity leap for people who can decompose problems cleanly and evaluate outputs critically. For a staff engineer, it's transformative. For a product manager who heard "you can just describe what you want in English," it's a confidence trap with a long fuse.

The line everyone will share: "that era is over." The line that actually matters: "it needs high-level direction, judgement, taste." The distance between those two sentences is where most of the real work still lives — and where it's going to live for a while yet.