Andrej Karpathy: From Vibe Coding to Agentic Engineering
Karpathy says we're past vibe coding — agentic engineering demands new taste and judgment.
Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, took the stage at Sequoia's AI Ascent 2026 alongside partner Stephanie Zhan. One year after coining the term 'vibe coding' to describe a casual, AI-assisted programming style, Karpathy declared the field has matured into something more rigorous: 'agentic engineering.' He described LLMs not as animals or tools but as 'ghosts' — jagged statistical entities that can be summoned but require unique taste and judgment to direct. Karpathy admitted he has never felt more behind as a programmer, emphasizing that the rapid evolution of AI agents demands deeper understanding, not just delegation.
Karpathy expanded on the concept of Software 3.0, where AI systems generate code from high-level specifications. He highlighted the critical limit of verifiability: when AI agents produce complex outputs, validating correctness becomes a human bottleneck. The key insight, he argued, is that you can outsource your thinking to AI but never your understanding. This distinction separates effective agentic engineers from casual vibe coders. For professionals, the message is clear: embracing AI agents requires developing a new kind of taste — knowing which tasks to delegate, when to intervene, and how to maintain comprehension of the system's behavior.
- Karpathy coined 'vibe coding' one year ago; now introduces 'agentic engineering' as the serious discipline for building with AI agents.
- He describes LLMs as 'ghosts' — jagged, statistical, summoned entities that require new taste and judgment to direct effectively.
- Warns that while thinking can be outsourced to AI, understanding cannot — verifiability remains a human bottleneck in Software 3.0.
Why It Matters
Karpathy's framework redefines professional AI collaboration: agents are powerful but demand human taste and understanding.