Vibe Coding: LLMs and Software Engineering Agents

The paper provides a comprehensive survey of "Vibe Coding," an emerging software development paradigm where developers rely on Large Language Model (LLM)-based autonomous coding agents and validate output through observation rather than line-by-line inspection. It establishes a formal definition of Vibe Coding as a triadic relationship between humans, software projects, and agents, formalised as a Constrained Markov Decision Process. The paper details the entire ecosystem, including the foundations of Code LLMs (data, pre-training, and post-training techniques), the architecture of coding agents (planning, memory, and reflection), and the essential development environment components (isolated execution runtimes and orchestration platforms). Finally, it synthesises existing practices into five distinct development models and discusses critical future challenges regarding scalability, security, and the evolving role of the human developer.