Abstract

The advancement of large language models (LLMs) has catalyzed a paradigmshift from code generation assistance to autonomous coding agents, enabling anovel development methodology termed "Vibe Coding" where developers validateAI-generated implementations through outcome observation rather thanline-by-line code comprehension. Despite its transformative potential, theeffectiveness of this emergent paradigm remains under-explored, with empiricalevidence revealing unexpected productivity losses and fundamental challenges inhuman-AI collaboration. To address this gap, this survey provides the firstcomprehensive and systematic review of Vibe Coding with large language models,establishing both theoretical foundations and practical frameworks for thistransformative development approach. Drawing from systematic analysis of over1000 research papers, we survey the entire vibe coding ecosystem, examiningcritical infrastructure components including LLMs for coding, LLM-based codingagent, development environment of coding agent, and feedback mechanisms. Wefirst introduce Vibe Coding as a formal discipline by formalizing it through aConstrained Markov Decision Process that captures the dynamic triadicrelationship among human developers, software projects, and coding agents.Building upon this theoretical foundation, we then synthesize existingpractices into five distinct development models: Unconstrained Automation,Iterative Conversational Collaboration, Planning-Driven, Test-Driven, andContext-Enhanced Models, thus providing the first comprehensive taxonomy inthis domain. Critically, our analysis reveals that successful Vibe Codingdepends not merely on agent capabilities but on systematic context engineering,well-established development environments, and human-agent collaborativedevelopment models.