Andrej Karpathy: From Vibe Coding to Agentic Engineering

Speaker: Andrej Karpathy (co-founder of OpenAI, former head of AI at Tesla, founder of Eureka Labs) Interviewer: Stephanie Zhan, Sequoia Capital Partner Collection: AI-Ascent-2026 Source: YouTube — Sequoia Capital

Summary

Andrej Karpathy, in conversation with Sequoia partner Stephanie Zhan at AI Ascent 2026, lays out a framework for understanding the shift from "vibe coding" (his own coinage from 2025) to what he now calls "agentic engineering" — a serious engineering discipline for coordinating fallible, stochastic, but extremely powerful AI agents without sacrificing professional software quality. He argues the speed-up for skilled practitioners exceeds 10x by a wide margin, but that the human remains the bottleneck for taste, judgment, spec design, and architectural understanding.

The conversation is anchored by Karpathy's Software 3.0 thesis: LLMs are not faster tools for writing code — they are a new computing paradigm where the context window replaces the source file as the primary programming surface. He illustrates this with two examples: (1) OpenClaw's installer, which is a prompt rather than a shell script, and (2) his own MenuGen app, which he built as a traditional web application only to realize the entire app is "spurious" — a single Gemini prompt with image generation achieves the same result with zero code. He extends this to LLM-generated knowledge bases as a category of software that could not have existed before because no deterministic code could transform a corpus of facts into a structured wiki.

Karpathy introduces a verifiability framework for predicting which domains AI will automate fastest: LLMs peak in capability where outputs can be verified (math, code) because those domains receive the most RL training investment. He describes LLMs as "jagged entities" — simultaneously capable of refactoring 100K-line codebases while recommending you walk 50 meters to a car wash instead of drive. He attributes this jaggedness to the intersection of what is verifiable and what labs choose to prioritize in training data. His advice to founders: find verifiable domains the labs haven't focused on yet and build your own RL environments. His parting insight: "You can outsource your thinking, but you can't outsource your understanding."