SKIP TO CONTENT
ON AIR — VIBE CODING ACADEMY · EN · LIVE
All articles
INDUSTRY ANALYSIS·March 7, 2026·9 MIN READ

Karpathy Says Vibe Coding Is Dead — Welcome to the Age of Agentic Engineering

By Alex Rivera

The Creator Kills His Creation

Andrej Karpathy — OpenAI co-founder, former Tesla AI director, and the person who coined "vibe coding" — has declared the concept passé. His replacement term: agentic engineering.

The distinction matters. Vibe coding was about feel — prompting an AI and trusting whatever came out. Agentic engineering is about structure — orchestrating AI agents under disciplined human oversight to produce production-grade software.

In Karpathy's framing: "You are not writing the code directly 99% of the time. You are orchestrating agents who do, and acting as oversight."

What Changed

The shift isn't philosophical — it's empirical. The data that forced the rethink:

  • AI co-authored code has 1.7x more major issues than human-written code in controlled studies
  • Open-source projects including Gentoo Linux and NetBSD have implemented outright bans on AI-generated submissions
  • 92% of US developers now use AI coding tools daily, but quality gates haven't kept pace
  • 41% of global code is now AI-generated, up from single digits in 2024

Vibe coding worked for prototypes and weekend projects. It doesn't work when 41% of production code is AI-generated and the defect rate is nearly double.

The Agentic Engineering Framework

Karpathy's framework has four principles that distinguish it from vibe coding:

1. Spec First, Prompt Second

Write a design document or specification before touching the AI. Define what you want, why you want it, and what "done" looks like. The AI gets the spec, not a vague instruction.

2. Break Work Into Tasks

Agentic engineering treats AI coding like project management. Instead of "build me an app," you create well-defined, testable tasks: "Create the authentication middleware with JWT validation, rate limiting, and session management."

3. Review Everything

Every line of AI-generated code gets the same scrutiny as a human pull request. No exceptions. The 52% of developers who skip review (per Sonar's 2026 survey) are building technical debt that compounds exponentially.

4. Testing Is the Differentiator

The single biggest difference between a vibe-coded prototype and a production application is the testing layer. Agentic engineering mandates automated tests for every AI-generated feature — unit, integration, and end-to-end.

Why MIT Disagrees (Sort Of)

Interestingly, MIT Technology Review named vibe coding one of the "10 Breakthrough Technologies of 2026" — the same year Karpathy declared it dead. This isn't a contradiction. MIT is recognizing the democratization impact: millions of non-developers can now build software. Karpathy is recognizing the quality gap: that software needs engineering discipline to be reliable.

Both are right. The breakthrough was making coding accessible. The evolution is making the results trustworthy.

The Market Validates Both

The numbers tell a nuanced story:

  • $4.7 billion: Current AI coding market, growing at 38% CAGR
  • 100,000+ projects per day: Built on Lovable alone
  • 75% of small startup developers: Use Claude Code as their primary tool
  • 56% of enterprise developers: Still prefer GitHub Copilot's guardrails

Startups vibe code to validate ideas. Enterprises need engineering discipline. Agentic engineering bridges both — the speed of AI generation with the rigor of structured oversight.

What This Means for You

If you're building with AI coding tools today, the transition from vibe coding to agentic engineering looks like this:

  1. Start with CLAUDE.md / cursor rules: Document your project's architectural decisions, coding standards, and constraints so AI agents follow them consistently
  2. Use agent orchestration: Tools like Claude Code's skills system and Cursor Automations let you define structured workflows rather than ad-hoc prompts
  3. Implement verification pipelines: Static analysis, security scanning, and architecture tests that run on every AI-generated commit
  4. Track AI contribution metrics: Know what percentage of your codebase is AI-generated and where the defect clusters are

The era of "just prompt it and ship it" produced remarkable demos and fragile products. The era of agentic engineering produces the same speed with the reliability that paying customers demand.

For a deep dive into building production-quality software with AI agents, the Vibe Coding Ebook covers the complete workflow — from spec writing to agent orchestration to automated verification — across 22 chapters with 200+ production-ready prompts.


Sources: The New Stack, Medium (Generative AI Revolution), Glide Blog, Addy Osmani, MIT Technology Review