vibe coding

«Andrej Karpathy» coined «vibe coding» in February 2025 to name prompting an AI and accepting the result with minimal verification. By his April 2026 interview he had already split the idea in two: «vibe coding» raises the floor, democratizing software development, while agentic engineering preserves the quality bar needed to coordinate spiky, stochastic agents. The author of the term had moved past it.

What follows in the sources is a fight over where the practice stops. Dave Farley calls AI-assisted development without continuous delivery a trap, a complexity bomb with a delayed fuse, and invokes the Jevons paradox: cheap code means more code, more integration points, more behavior to evaluate. Simon Wardley reaches for the same paradox and a figure, a large company going from 30 million to over a billion lines of code just to hold its position. Olivier Rafal judges «vibe coding» untenable in the enterprise. Philippe Martin offers BMAD as the urban plan that answers it.

The scoping that survives is narrow: prototypes, MVPs, scripts, throwaway projects, and turning UX uncertainty into specifications. Hugo Lassiège, whose production code is nearly 100% generated, refuses the label and calls his practice «Context Engineering». Andrew Ng locates the failure precisely: the inexperienced vibe coder does not know what context to give his agent.

Against all of that, «Gene Kim» reports «Dario Amodei»'s definition, an iterative conversation that results in AI writing your code, and says that at Anthropic there is no other game in town.