vibe coding

Andrej Karpathy coined the term in February 2025 for a narrow loop: prompt a model and accept the result with minimal verification. He also drew the line most later writers inherit: vibe coding raises the floor by democratising software development, while agentic engineering preserves the quality bar by disciplining how agents are coordinated. Its stated home is disposable work, prototypes, MVPs, scripts, learning.

The recurring question is where that home ends. Donn Felker reports a CFO stuck after two months inside a single Claude Cowork session with no idea where their code lived, and a photographer whose Base44 application stored media on Google Drive at an untenable transfer cost. Session persistence, code location and infrastructure choices are the places the method gives way.

Dave Farley classes it as an anti-pattern: a vague spec delegated to an AI with no deployment pipeline, a complexity bomb with a delayed fuse. Simon Wardley applies the Jevons paradox and projects a large enterprise going from 30 million to over a billion lines of code. Addy Osmani and his co-authors drop the binary and place the practice on a spectrum of structure, verification and judgment, where its low CapEx and high OpEx finish three to ten times more expensive per feature than agentic engineering. Andrew Ng puts the failure in the practitioner: the inexperienced vibe coder does not know what context to give their agent. Hugo Lassiège, whose production code is almost entirely generated, refuses the label and calls his practice Context Engineering instead.