"AI writes 80% of my code": the author of "vibe coding" changed his view on AI agents in just 3 months

In a detailed post, Karpathy shared that as recently as November 2025 he was writing about 80% of his code manually and using AI only occasionally. By December, the proportion had flipped: AI agents handled approximately 80% of the work, while he primarily made edits.

«I now mostly program in English, somewhat awkwardly explaining to the LLM what code needs to be written… with words,» Karpathy admitted. He acknowledged that «this hits the ego,» but said the ability to edit large code segments at once was too useful to abandon.

A few random notes from claude coding quite a bit last few weeks.

Coding workflow. Given the latest lift in LLM coding capability, like many others I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in…

— Andrej Karpathy (@karpathy) January 26, 2026

He describes this shift as the biggest change in his working practice in the last 20 years, suggesting that a «double-digit percentage of engineers» are experiencing a similar transition, though it remains largely unnoticed by the general public.

Karpathy emphasized that AI agents are far from perfect. He compared their errors to those of «a slightly careless and rushed junior developer»: models can make incorrect assumptions, build solutions on them, fail to ask clarifying questions, not highlight contradictions, and often overcomplicate code. «If the code really matters, you need to watch them like a hawk,» he warns.

He specifically criticized AI’s tendency to inflate abstractions: «They will easily produce a thousand lines where everything could be reduced to a hundred if you nudge them a bit.»

Karpathy warned of a «slopocalypse» — an avalanche of nearly correct but low-quality AI-generated content on GitHub, in scientific publications, social networks, and media. Nevertheless, he acknowledged that the transition to agent-based programming is «a huge net positive»: the main advantage is not merely speed but the ability to take on projects that were previously too complex or impractical due to knowledge gaps.

Karpathy also noted that AI coding might divide engineers into two camps: those who enjoy writing code and those who prefer focusing on building products with AI. He admitted his manual programming skills were gradually atrophying: «The ability to review code is not the same as the ability to write it.»