What Is Vibe Coding? Origin, Tools, and the Technical Debt Reality (2026)

Vibe coding means prompting AI to write code without reviewing it. Coined by Andrej Karpathy in February 2025, named Collins Word of the Year. Here's what it actually means, the tools behind it, and why it's controversial.

March 1, 2026 · 3 min read

Vibe coding means prompting an AI to write code and shipping it without reading the output. Andrej Karpathy coined the term in February 2025. Collins Dictionary named it Word of the Year. One year later, the industry is reckoning with the consequences.

What Does Vibe Coding Mean?

Vibe coding is a software development practice where you describe what you want in natural language and accept AI-generated code without fully reviewing it. You prompt, you run it, and if it works, you move on. The code itself is treated as an implementation detail you don't need to understand.

The defining characteristic is acceptance without comprehension. Simon Willison drew the clearest line: if an LLM wrote the code but you reviewed it, tested it, and can explain how it works, that's not vibe coding. That's software development. Vibe coding specifically means shipping code you haven't read.

92%
US devs using AI coding tools daily
41%
Global code that is AI-generated
$4.7B
Vibe coding tools market (2026)
75%
Replit users who never write code

Who Coined Vibe Coding? The Origin Story

Andrej Karpathy, co-founder of OpenAI and former AI leader at Tesla, posted on X on February 2, 2025:

Karpathy's original post

"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even type."

The post was viewed over 4.5 million times. Karpathy later called it a "shower thought throwaway tweet" that he "just fired off," but it named something millions of developers were already doing. The term spread fast:

  • March 2025: Merriam-Webster listed "vibe coding" as a slang and trending term
  • November 2025: Collins Dictionary named it Word of the Year for 2025
  • February 2026: One-year anniversary with widespread retrospectives across the industry

At the time, Karpathy was using Cursor Composer with Anthropic's Claude Sonnet model and SuperWhisper for voice-to-text. He described building entire projects by talking to his computer, seeing error messages, and pasting them back into the chat without understanding the underlying code.

Vibe Coding Tools: What People Actually Use

The tools split into two categories: app builders that generate complete applications from conversation alone, and code editors that integrate AI into professional development workflows.

ToolTypeKey StatVibe Coding Fit
LovableApp builder$100M ARR in 8 monthsPure vibe coding: describe app, get app
Bolt.newApp builderFastest to prototype (28 min avg)Prompt to full-stack app in browser
Replit AgentApp builder$10M to $100M revenue in 9 months75% of users never write code
v0 by VercelApp builderReact/Next.js focusedStrong for UI prototyping
CursorCode editor360K paying users, $9.9B valuationSupports vibe coding but also engineering workflows
WindsurfCode editor#1 LogRocket rankingsArena Mode, 5 parallel agents
Claude CodeTerminal agent$2.5B ARRAgentic, not vibe: designed for reviewed code
GitHub CopilotCode editor15M+ developersInline completions, chat, and agent mode

App builders vs. code editors

App builders (Lovable, Bolt, Replit, v0) are where pure vibe coding happens: you describe what you want and get a working app. Code editors (Cursor, Windsurf, Copilot) support vibe coding as one workflow but also enable careful, reviewed development. AI coding agents like Claude Code and Codex CLI are designed for agentic workflows where code is reviewed before shipping.

When Vibe Coding Works (and When It Doesn't)

Prototyping and MVPs

Vibe coding's best use case. Go from idea to working demo in minutes. Bolt.new averages 28 minutes to a working prototype. Ideal for validating ideas before investing engineering time.

Learning and Exploration

Non-programmers can build functional tools to understand what's possible. Developers can explore unfamiliar frameworks quickly. The low barrier to entry is genuinely valuable for education.

Internal Tools and Scripts

One-off scripts, data transformations, and internal dashboards where the blast radius is small. If it breaks, you regenerate it. Low stakes make the lack of review acceptable.

Weekend Projects

This was Karpathy's original context. Personal projects where you're the only user, the code doesn't handle sensitive data, and technical debt is irrelevant because there's no team to inherit it.

Where Vibe Coding Fails

The problems start when vibe-coded projects graduate from prototype to production:

  • Security: 45% of AI-generated code contains security vulnerabilities. A December 2025 study found 69 vulnerabilities across 15 applications built with vibe coding tools. Every tool tested introduced Server-Side Request Forgery (SSRF) flaws.
  • Maintenance: Code you don't understand is code you can't debug. When something breaks at 2 AM, the LLM that wrote it doesn't remember the context.
  • Scale: AI-generated code duplication increased 48% while refactoring activity dropped 60%. The codebase grows faster than anyone can comprehend it.
  • Cost: A leaked Supabase database from a vibe-coded app exposed 1.5 million API keys and 35,000 user emails. The cost of a breach dwarfs the time saved.

The Technical Debt Problem

Multiple independent studies converge on the same conclusion: vibe coding accelerates code production while degrading code quality at a pace that compounds faster than teams can address.

45%
AI-generated code with security flaws
48%
Increase in code duplication
60%
Drop in refactoring activity
69
Vulnerabilities found in 15 vibe-coded apps

The workforce angle makes this worse. Junior developers learn by writing code, making mistakes, and debugging them. If companies hire fewer juniors because AI handles their tasks, the engineers needed to fix 2026's technical debt in 2028 won't exist. Multiple analysts are calling 2026-2027 the peak technical debt era driven by vibe coding.

The counterargument: AI tools will keep improving and eventually fix their own debt. But that assumes the tooling improves faster than the debt accumulates, which hasn't been the pattern so far.

The Open-Source Crisis

The open-source impact has been severe enough that maintainers are taking drastic action. The flood of AI-generated contributions, sometimes called "AI Slopageddon," hit a breaking point in early 2026:

  • Daniel Stenberg shut down cURL's six-year bug bounty program because of AI-generated spam reports
  • Mitchell Hashimoto banned AI-generated code from Ghostty
  • Steve Ruiz had tldraw auto-close all external pull requests

The structural threat goes deeper. Vibe coding tools install open-source packages without developers reading documentation, reporting bugs, or engaging with maintainers. Tailwind CSS downloads climbed while documentation traffic fell 40% and revenue dropped 80%. Stack Overflow saw 25% less activity within six months of ChatGPT's launch.

Open-source sustainability depends on a feedback loop: users find bugs, report them, contribute fixes, and sometimes pay for enterprise features. Vibe coding breaks every link in that chain by inserting an AI layer between developers and the projects they depend on.

Vibe Coding vs. Agentic Coding

These terms get conflated, but they describe fundamentally different relationships between humans and AI-generated code.

DimensionVibe CodingAgentic Coding
Developer roleDescribes what they wantArchitects, reviews, and directs agents
Code reviewNone or minimalThorough review before shipping
Code understandingNot requiredRequired for production use
AutonomyHuman prompts each stepAgent plans and executes multi-step tasks
Best forPrototypes, MVPs, learningProduction code, large codebases, enterprise
ToolsLovable, Bolt, Replit, v0Claude Code, Codex CLI, Devin, Cursor agent mode
Risk profileHigh (unreviewed code)Managed (human oversight on outputs)

An academic paper (May 2025) formalized the distinction: vibe coding emphasizes intuitive, prompt-based workflows for experimentation, while agentic coding enables autonomous development through goal-driven agents that plan, execute, test, and iterate.

In practice, professional developers increasingly use agentic tools while maintaining engineering rigor. The industry is moving from "accept everything" toward "direct agents, review outputs." Karpathy himself now calls this approach agentic engineering.

Where Karpathy Landed

One year after coining the term, Karpathy's own practice has evolved. His latest project, Nanochat (a minimal ChatGPT-like web UI), was hand-written. He tried using Claude and Codex agents but found them "just didn't work well enough at all and net unhelpful, possibly the repo is too far off the data distribution."

His preferred term is now agentic engineering: "agentic because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight. Engineering to emphasize that there is an art and science and expertise to it."

This tracks with the broader industry trajectory. Vibe coding named a real phenomenon, but the long-term direction is toward AI-assisted engineering with human oversight, not blind acceptance of AI outputs. The tools are getting better, but the need for human judgment on architecture, security, and maintainability isn't going away.

Frequently Asked Questions

What is vibe coding?

Vibe coding is a software development practice where you describe what you want in natural language and accept AI-generated code without fully reviewing or understanding it. The term was coined by Andrej Karpathy on February 2, 2025. Collins Dictionary named it Word of the Year for 2025.

Who coined the term vibe coding?

Andrej Karpathy, co-founder of OpenAI and former AI leader at Tesla. He posted about it on X on February 2, 2025, describing it as "fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists." The post was viewed over 4.5 million times.

Is vibe coding bad?

It depends on context. For prototypes, weekend projects, and learning, it's effective. For production systems, it's risky: 45% of AI-generated code contains security vulnerabilities, code duplication increased 48%, and the open-source community has been hit by AI-generated spam contributions. The term's own creator now prefers "agentic engineering" for serious work.

What is the difference between vibe coding and agentic coding?

Vibe coding means describing what you want and accepting AI output without deep review. Agentic coding means setting a goal and letting an autonomous agent plan, write, test, and iterate, with human oversight on the results. Professional developers use agentic tools while reviewing code before it ships.

What tools are used for vibe coding?

App builders (Lovable, Bolt.new, Replit Agent, v0) are the purest vibe coding tools. Code editors (Cursor, Windsurf, Copilot) support vibe coding as one workflow among many. Terminal agents (Claude Code, Codex CLI) are designed more for agentic coding than pure vibe coding.

Does Andrej Karpathy still use vibe coding?

He has moved past it. His latest project was hand-written because AI agents "didn't work well enough." He now prefers "agentic engineering," which emphasizes orchestrating AI agents while maintaining engineering standards. He described the original vibe coding tweet as a "shower thought throwaway" that named something many developers were feeling.

Build with AI, Ship with Confidence

Whether you're vibe coding a prototype or engineering production systems, every AI tool needs a reliable apply layer. Morph's Fast Apply model merges LLM edits deterministically at 10,500+ tokens per second.