In February 2025, computer scientist and AI researcher Andrej Karpathy coined the phrase "vibe coding" to describe an emerging approach to software development. Instead of manually writing syntax line-by-line, developers use natural language prompts to guide AI models in generating, editing, and running software prototypes.
🛠️ Hands-On Reality Check
During our internal testing with Cursor (powered by Claude 3.7 Sonnet), we prototyped a full-stack dashboard in under 15 minutes. While the initial generation speed is impressive, setting up complex backend state management still required manual human oversight and debugging—highlighting why "vibes" alone aren't enough for production-grade software.
What Exactly Is Vibe Coding?
Vibe coding is an application development method where you describe your idea in natural language, and an AI tool generates the code. The process involves repeatedly instructing the tool on desired changes until you reach the intended result. It allows creators to focus on the high-level concept without micromanaging every line of code.
Editing involves instructing the tool after each output until the feature works as intended. For experienced developers, it can also include manually refining or replacing generated code where necessary.
Traditional Coding vs. Vibe Coding
| Feature | Traditional Coding | Vibe Coding |
|---|---|---|
| Primary Input | Manual lines of code, precise syntax | Plain natural language instructions |
| Developer Focus | Detailed implementation & syntax | High-level architecture & product vision |
| Debugging Method | Reading stack traces line-by-line | Iterative natural language prompting & AI fixes |
| Skill Requirement | Programming knowledge is usually important | Lower initial barrier, but technical knowledge remains valuable for complex projects |
Recommended Vibe-Coding Tech Stack:
- Primary Editors: Cursor, Windsurf (Agentic Mode)
- Top LLMs: Claude 3.7 Sonnet, OpenAI o3-mini
- Security Guard: Socket.dev (To detect AI-hallucinated packages)
Who Is Vibe Coding For?
Vibe coding is for software developers, graphic designers, entrepreneurs, product managers, and anyone with digital ideas who wants to build software efficiently.
Why Vibe Coding in 2026?
Vibe coding has gained significant attention because it streamlines problem-solving for both engineers and non-technical creators. Tedious coding tasks are reduced to instruction-based prompts, yielding functional outputs based on how precise the directions are. Users can iteratively refine prompts until achieving the desired result.
The underlying models and developer interfaces have matured significantly, moving past experimental tech toward reliable software assembly.
According to Y Combinator managing partner Jared Friedman, about 25% of the Winter 2025 batch had codebases in which 95% of the code was generated by AI. Friedman noted that these founders were technically capable of building their products themselves, but AI was handling much of the implementation.
The trend has only accelerated since. Stack Overflow's 2025 Developer Survey found that 84% of developers worldwide now use or plan to use AI coding tools, up from 76% the year before, though only around 29-33% say they fully trust the accuracy of that generated code. GitHub reports that AI now writes over 40% of new code globally, and Gartner projects that 90% of enterprise software engineers will be using AI code assistants by 2028. Adoption is especially strong across Asia: India alone accounts for roughly 16-17% of the global vibe coding user base, with APAC as a region driving close to 40% of overall usage.
Figure: Vibe coding adoption metrics for 2026, based on the Stack Overflow Developer Survey, GitHub, and Gartner.
Vibe coding still requires clear thinking. To achieve strong results, you must articulate the core vision — defining the problem being solved, the target user, and what successful implementation looks like. The AI tool does not replace human domain knowledge or logic; it accelerates execution based on your direction.
From Vibe Coding to Agentic Engineering: Karpathy's Own Shift
The term has kept evolving even in the mouth of the person who coined it. In December 2025, Karpathy said his own workflow flipped from roughly 80% hand-written code to about 80% agent-generated code within a few weeks, a change he described as a "phase transition" rather than a gradual shift. By February 2026, in a post he later expanded on stage at Sequoia Capital's AI Ascent event that April, Karpathy said vibe coding — the loose, low-scrutiny style he had described a year earlier — was already "passé," and proposed a more disciplined successor he calls "agentic engineering."
The distinction he draws: vibe coding raises the floor, letting anyone with an idea prototype something that mostly works. Agentic engineering is meant to preserve the ceiling — professional-grade software built by directing, reviewing, and constraining AI agents with real oversight, rather than accepting the first draft on vibes alone.
Vibe Coding by the Numbers
| Metric | Figure | Source |
|---|---|---|
| Developers who use or plan to use AI coding tools | 84% (up from 76%) | Stack Overflow 2025 Developer Survey |
| Developers who trust the accuracy of that output | 29% (down from 40%) | Stack Overflow 2025 Developer Survey |
| Enterprise software engineers expected to use AI code assistants by 2028 | 90% (up from <14% in early 2024) | Gartner |
| YC Winter 2025 batch with 95%+ AI-generated codebases | 25% | Y Combinator (Jared Friedman) |
| AI-generated code samples introducing an OWASP Top 10 vulnerability | 45% | Veracode, 2025 GenAI Code Security Report |
| More major issues / security vulnerabilities in AI-co-authored PRs vs. human-only code | 1.7x / 2.74x | CodeRabbit |
| Global vibe coding usage share held by India / APAC | 16-17% / ~40% | Vercel |
AI Tools for Vibe Coding
Several AI tools allow you to build applications using natural language instructions. Each has distinct strengths depending on your project scope:
| Tool Name | Key Capabilities | Target Audience |
|---|---|---|
| Lovable / Bolt.new | Generates full-stack web apps and MVPs from simple text prompts. | Beginners, founders, and non-technical creators. |
| Claude Code | Command-line tool for multi-file editing, complex tasks, and agentic workflows. | Engineers managing large codebases. |
| Replit | Browser-based interactive environment displaying code in real time. | Learners and interactive developers. |
| Kimi | Rapid generation of web prototypes and generative AI design concepts. | Fast visual prototypers. |
| Cursor | AI-first code editor with full-codebase context awareness and fast inline editing. | Developers building custom software. |
| v0 by Vercel | Generates clean front-end UI layouts, landing pages, and components. | UI/UX designers and front-end developers. |
After setting up an initial prototype, you can visually refine layouts, adjust content, and publish the finished application.
Pros and Cons of Vibe Coding
Pros
- Speed and Efficiency: Developers can focus on high-level architecture, user experience, and product vision while delegating repetitive boilerplate code.
- Rapid Prototyping: Concepts can be quickly converted into functional prototypes, allowing teams to test ideas early.
- Accessibility: Reduces technical barriers, enabling founders and designers without deep programming backgrounds to build software.
- Creative Focus: Encourages clear conceptual thinking and problem definition, shifting energy from syntax errors to product design.
Cons
- Technical Oversights: Over-reliance on AI can lead to architectural flaws or unoptimized code if not properly audited.
- Security and Authentication Risks: AI-generated code may contain security vulnerabilities, flawed authentication logic, or unhandled edge cases. Veracode's 2025 GenAI Code Security Report, which tested more than 100 large language models across 80 coding tasks, found that 45% of AI-generated code samples introduced a known OWASP Top 10 vulnerability — a rate that stayed essentially flat in its Spring 2026 retest of newer flagship models. Technical oversight is necessary to ensure data privacy and prevent security breaches.
- Debugging Complexity: As Karpathy noted, the final phase of refining a complex project can still present challenges. Debugging stack traces and diagnosing edge-case bugs often require traditional technical knowledge.
Actionable Guardrail: To prevent AI tools from reading or indexing sensitive credentials, create a .cursorignore file in your project root:
Conclusion
Vibe coding in 2026 represents a pragmatic shift in how software is created. By pairing natural language instructions with AI execution, both technical developers and non-technical creators can bring ideas to life faster than ever. However, building robust software still requires human oversight, security auditing, and critical thinking to turn AI-generated drafts into secure, production-ready applications. Even Karpathy, who coined the term, now argues that professional teams need to move past "vibes" toward the more supervised discipline of agentic engineering.
Frequently Asked Questions
Common questions on vibe coding and AI-assisted software development in 2026.
Vibe coding is a style of software development, coined by Andrej Karpathy in February 2025, in which a developer describes what they want in natural language and an AI tool generates, runs, and refines the code, letting the person focus on the high-level idea rather than the syntax.
Not quite. AI pair programming still has a developer writing and reviewing most of the code with AI suggestions. Vibe coding shifts further, letting the AI generate the majority of the implementation while the person focuses on describing outcomes and testing results.
It can be, but only with proper review. Veracode's testing of over 100 large language models found that 45% of AI-generated code samples introduced a security vulnerability, so production-bound projects still need security auditing, testing, and technically capable oversight before launch.
Non-technical creators typically start with Lovable, Bolt.new, or v0, since these tools generate full working prototypes from a simple text prompt with minimal setup. Developers who want more control over an existing codebase tend to prefer Cursor or Claude Code.
In a February 2026 post, later expanded on stage at Sequoia's AI Ascent event that April, Karpathy said AI agents had gotten capable enough that professionals were shifting to a more supervised workflow he calls "agentic engineering," which keeps the speed of vibe coding but adds far more review and scrutiny before code ships.
Most industry voices, including Y Combinator's leadership, argue it won't. Vibe coding speeds up implementation, but reading code, catching bugs, and making architectural decisions at scale still requires human technical judgment that current AI models can't fully replace.
Sources
Stack Overflow, 2025 Developer Survey
Gartner, Top Strategic Trends in Software Engineering for 2025 and Beyond
Andrej Karpathy, Sequoia AI Ascent 2026 session notes
