Karpathy, Agents Redefine Vibe Coding

The sharpest takeaway is that AI coding is maturing from improvisation into operations: less typing, more orchestration. Karpathy is basically saying the hard part is no longer getting code out of the model, but directing unreliable systems without losing your own understanding.

• –Vibe coding works for rough prototypes; agentic engineering is about coordinating tools, context, and review across multi-step work
• –His “ghosts” framing is the right mental model: LLMs are powerful, but not deterministic enough to treat like conventional software
• –The limits of verifiability matter more as agents take on bigger tasks, because trust shifts from code inspection to process design
• –Software 3.0 remains the big picture, but the winning skill is increasingly judgment, not raw implementation speed
• –For developers, this is a category shift: the bottleneck moves from syntax to supervision, evals, and taste