Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan

Summary

Andrej Karpathy talks with Stephanie Zhan about how AI coding agents crossed a capability threshold around December, shifting from 'vibe coding' to a more disciplined 'agentic engineering.' He explains Software 3.0, where prompts and context replace traditional code and app logic, and why LLMs are jagged, verifiable-domain-optimized 'ghosts.' He also discusses agent-native infrastructure, the need for human taste and understanding, and implications for education.

• Karpathy says he has never felt more behind as a programmer after agentic tools improved sharply around December.
• He frames Software 3.0 as programming through prompts and context, with examples like OpenClaw installation and MenuGen vs Gemini/Nano Banana.
• Verifiability helps explain why models excel in math and code but remain jagged elsewhere.
• Agentic engineering aims to preserve professional software quality while using fallible, stochastic agents.
• Human taste, judgment, and understanding remain key bottlenecks as agents handle more implementation.
• He expects more agent-native infrastructure and agent representation for people and organizations.
• Education should focus on understanding, since thinking can be outsourced but understanding cannot.