---
id: "lib-2026-04-29-andrej-karpathy-from-vibe-coding-to-agentic"
type: "library"
title: "Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan"
lang: "EN"
org: "Sequoia Capital"
source_type: "Talk"
date: "2026-04-29"
week: "2026-W18"
url: "https://nosov.ai/library/lib-2026-04-29-andrej-karpathy-from-vibe-coding-to-agentic/"
original_url: "https://www.youtube.com/watch?v=96jN2OCOfLs"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Andrej Karpathy, OpenAI co-founder and former Tesla AI lead, in a 30-minute Sequoia fireside; sharp framing (agentic engineering, verifiability, jagged models) from one of the field's most-followed practitioners, but anecdotal and largely a recap of his own earlier essays."
summary_basis: "transcript"
original_minutes: "30"
tags: ["ai technology", "lecture"]
---

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

Tier B (Useful, limited): Andrej Karpathy, OpenAI co-founder and former Tesla AI lead, in a 30-minute Sequoia fireside; sharp framing (agentic engineering, verifiability, jagged models) from one of the field's most-followed practitioners, but anecdotal and largely a recap of his own earlier essays.

Summary based on the full transcript.

## In short

Andrej Karpathy talks with Sequoia's Stephanie Zhan (April 2026) about how his own programming changed after models improved in December 2025. He separates "vibe coding", which raises the floor so anyone can build software, from "agentic engineering", which keeps professional quality while going faster. Models are "jagged": strong where outputs can be verified and labs built training environments, weak elsewhere, so humans stay in charge of spec, taste and oversight.

## Key numbers

- December: month (2025) Karpathy names as the point when agent-written code chunks stopped needing correction
- more than 10x: speed-up he sees for the best agentic engineers, above the old '10x engineer' notion
- 50 m: car-wash distance in the question frontier models still answer wrongly ('walk'), his example of jagged capability

## What it reports

- Software 3.0: programming becomes prompting, and the context window is the lever. Example: OpenClaw installs via text pasted to an agent instead of a shell script.
- His MenuGen app (photo of a menu, images of dishes) is made redundant by asking Gemini to render images onto the photo directly. Lesson: look for things that need no app, not only faster versions of old apps.
- LLM-built knowledge bases are a new kind of information processing that had no code equivalent.
- Verifiability: classical computers automate what can be specified; LLMs automate what can be verified, because labs train with reinforcement learning on verifiable rewards. Capability also depends on what labs choose to put in the data, such as chess data added between GPT-3.5 and GPT-4.
- Advice to founders: verifiable domains outside lab focus can be tackled with your own RL environments and fine-tuning.
- Hiring for agentic engineers should use large projects (for example, build and secure a Twitter clone that another agent then tries to break), not puzzles.
- Agent mistakes persist: his agent matched Stripe and Google accounts by email address instead of a persistent user ID. Generated code is often bloated with brittle abstractions, and models resist simplification.
- Documentation is still written for humans; he wants text to paste to an agent, and deployment (DNS, service settings) as a test of agent-native infrastructure.
- On learning: "you can outsource your thinking, but you can't outsource your understanding" (a tweet he quotes).

## How they got it

Conversation drawing on Karpathy's personal projects and his previous writing on verifiability and "animals vs ghosts". No data or systematic comparison.

## Limits

- Personal experience of an elite programmer working on side projects, not team or enterprise settings.
- The more-than-10x claim is an impression, not a measurement.
- Auto-captioned; product names (OpenClaw, Opus 4.7, Codex 5.4) appear as spoken.

## Why it matters for companies

The verifiability lens is a practical way to rank where agents will work first: tasks with a checkable result (tests pass, numbers reconcile) before tasks judged on taste. The vibe coding vs agentic engineering split is useful for policy: let staff build freely for themselves, but hold anything shipped to others to the old quality bar. Hiring and assessment for technical roles should test how candidates direct and verify agents on real projects.

Original: https://www.youtube.com/watch?v=96jN2OCOfLs