Skip to content

Is Andrej Karpathy wrong about agentic coding?

A recent clip from his interview with Dwarkesh Patel has been doing the rounds. In it, he calls current coding agents “not net useful” for his nanochat repo – and a lot of people have since quietly updated their beliefs about what’s possible.

1 min read

Originally posted on LinkedIn · 35 reactions · 12 comments · View original →

Is Andrej Karpathy wrong about agentic coding? 👇

A recent clip from his interview with Dwarkesh Patel has been doing the rounds. In it, he calls current coding agents “not net useful” for his nanochat repo – and a lot of people have since quietly updated their beliefs about what’s possible.

I think that’s the wrong generalisation.

In this long-form piece, I unpack:

  • the very peculiar context he’s speaking from (nanochat + his craft),
  • why vibe coding and serious AI Engineering are not the same thing,
  • Human + Agent Fleet vs Spec + Agent Swarm
  • What this all means for real teams trying to ship software with agents today.

Karpathy is right about his own context. He’s incomplete as a spokesperson for the whole industry.

Please, share your thoughts: how are you actually using (or avoiding) agentic coding in your team today?