A Survey of Vibe Coding with Large Language Models (2510.12399v1)

Abstract: The advancement of LLMs has catalyzed a paradigm shift from code generation assistance to autonomous coding agents, enabling a novel development methodology termed "Vibe Coding" where developers validate AI-generated implementations through outcome observation rather than line-by-line code comprehension. Despite its transformative potential, the effectiveness of this emergent paradigm remains under-explored, with empirical evidence revealing unexpected productivity losses and fundamental challenges in human-AI collaboration. To address this gap, this survey provides the first comprehensive and systematic review of Vibe Coding with LLMs, establishing both theoretical foundations and practical frameworks for this transformative development approach. Drawing from systematic analysis of over 1000 research papers, we survey the entire vibe coding ecosystem, examining critical infrastructure components including LLMs for coding, LLM-based coding agent, development environment of coding agent, and feedback mechanisms. We first introduce Vibe Coding as a formal discipline by formalizing it through a Constrained Markov Decision Process that captures the dynamic triadic relationship among human developers, software projects, and coding agents. Building upon this theoretical foundation, we then synthesize existing practices into five distinct development models: Unconstrained Automation, Iterative Conversational Collaboration, Planning-Driven, Test-Driven, and Context-Enhanced Models, thus providing the first comprehensive taxonomy in this domain. Critically, our analysis reveals that successful Vibe Coding depends not merely on agent capabilities but on systematic context engineering, well-established development environments, and human-agent collaborative development models.

• The paper formalizes Vibe Coding by modeling the collaboration of humans, coding agents, and projects as a constrained Markov Decision Process.
• It synthesizes over 1000 research works to develop a detailed taxonomy covering data foundations, training strategies, and coding agent architectures.
• It identifies open challenges in context engineering, security, and scalable oversight, setting a roadmap for future AI-driven software development.

A Survey of Vibe Coding with LLMs: Technical Foundations, Development Models, and Open Challenges

Introduction and Motivation

The surveyed paper provides a comprehensive and systematic review of "Vibe Coding," a paradigm shift in software engineering catalyzed by the rapid advancement of LLMs and autonomous coding agents. Vibe Coding is characterized by a triadic collaboration among human developers, coding agents, and software projects, where developers validate AI-generated implementations through outcome observation and iterative feedback, rather than traditional line-by-line code comprehension. This paradigm is formalized as a Constrained Markov Decision Process (MDP), capturing the dynamic interplay between human intent, project context, and agentic action.

The paper synthesizes over 1000 research works, establishing theoretical foundations, practical frameworks, and a taxonomy of development models for Vibe Coding. It further identifies critical challenges in context engineering, infrastructure, security, and human-AI collaboration, providing a technical roadmap for future research and deployment.

Figure 1: Vibe Coding is a triadic collaboration among developer, coding agent, and project, with iterative instruction–feedback loops enabling context-aware coding and result-oriented review.

Theoretical Formalization of Vibe Coding

Vibe Coding is defined as an engineering methodology where the human ($\mathcal{H}$), project ($\mathcal{P}$), and coding agent ($\mathcal{A}_\theta$) interact in a closed-loop system. The human articulates requirements and evaluates outputs, the project provides the contextual information space, and the agent executes conditional code generation and modification. The collaboration is modeled as a Constrained MDP:

$$\mathcal{V}_{\text{MDP}} = \langle \mathcal{S}_{\mathcal{P}}, \mathcal{A}_{\mathcal{H} \rightarrow \mathcal{A}_\theta}, \mathcal{T}_{\mathcal{A}_\theta | \mathcal{P}}, \mathcal{R}_{\mathcal{H}}, \gamma \rangle$$

where the state space is defined by the project, actions are triggered by human instructions, transitions are constrained by project specifications, and rewards are determined by human evaluation.

The agent's code generation is contextually conditioned on human intent, project context, and execution environment, with the optimization objective being the orchestration of context to maximize alignment with human expectations under context window constraints.

Iterative task expansion is formalized, supporting dynamic requirement evolution and progressive constraint reinforcement, distinguishing Vibe Coding from traditional paradigms with frozen requirements.

Figure 2: The evolution timeline of Vibe Coding, showing the progression from foundational coding LLMs to sophisticated agent systems, execution environments, and feedback mechanisms.

LLMs for Coding: Data, Pre-training, and Post-training

The survey provides a detailed taxonomy of LLMs for code, covering data foundations, pre-training, and post-training:

• Data Foundations: Training corpora are sourced from open repositories (e.g., GitHub, Stack Overflow), with both depth-focused (quality in popular languages) and breadth-focused (coverage across languages) strategies. Instruction datasets are constructed from real-world code, synthetic instructions, and preference data, with advanced pipelines for filtering, deduplication, and quality enhancement.
• Pre-training: Objectives include masked language modeling, autoregressive modeling, denoising, structure-aware tasks (e.g., data flow prediction), and contrastive learning. Multimodal and cross-modal pre-training further enhance code understanding and generation.
• Continual Pre-training: CPT enables domain adaptation while mitigating catastrophic forgetting via data replay and optimal data mixing.
• Post-training: Supervised fine-tuning (SFT) and instruction tuning align models with human intent, with parameter-efficient methods (e.g., LoRA, adapters) reducing computational overhead. Reinforcement learning (RLHF, DPO, GRPO) and verifiable reward mechanisms further refine reasoning and correctness, especially in code generation.

  Figure 3: The LLM training lifecycle for coding, encompassing data foundation, pre-training, and post-training, with core capabilities in code generation, understanding, bug detection, and optimization.

Coding Agent Architectures

The paper surveys the architecture of LLM-based coding agents, decomposing them into core components:

• Planning and Decomposition: Agents employ chain-of-thought, tree-of-thought, and hierarchical planning to decompose complex tasks, leveraging both prompting and external planning modules.
• Memory Mechanisms: Short-term memory is managed via context windows; long-term memory is implemented through external vector stores, memory banks, and dual-memory architectures, supporting retrieval, consolidation, and belief updating.
• Action Execution: Tool integration enables agents to invoke compilers, debuggers, and external APIs. Code-based action frameworks unify diverse behaviors, with multi-agent and role-based systems supporting complex workflows.
• Reflection and Debugging: Iterative refinement, self-reflection, and multi-agent critique loops enable agents to learn from errors, validate outputs, and autonomously debug code.
• Collaboration: Multi-agent systems distribute roles (e.g., programmer, tester, critic), coordinate via shared message pools, and employ structured communication protocols for scalable, resilient development.

  Figure 4: Coding agent architecture, highlighting cognitive system, memory, and tool integration, with planning, action execution, and multi-agent collaboration.

Development Environments for Coding Agents

The infrastructure for Vibe Coding is categorized into:

• Isolated Execution Environments: Containerization (Docker, LXC), sandboxing, and hardware-assisted isolation ensure secure, reproducible code execution. Cloud-based platforms orchestrate large-scale, distributed agent workloads.
• Interactive Development Interfaces: AI-native IDEs integrate inline completion, conversational agents, and contextual memory. Protocol standards (MCP, LSP, DAP) enable seamless tool integration and context exchange.
• Distributed Orchestration Platforms: CI/CD pipelines, cloud compute orchestration, and multi-agent frameworks (AutoGen, CrewAI, MetaGPT) support scalable, automated, and collaborative agentic workflows.

  Figure 5: Development environment architecture for coding agents, including isolated, interactive, and distributed platforms.

Feedback Mechanisms

Feedback is central to Vibe Coding, with a taxonomy spanning:

• Compiler Feedback: Syntax/type error detection, static analysis, and runtime compilation feedback are integrated into agent workflows for iterative refinement.
• Execution Feedback: Unit and integration test results, runtime error handling, and exception management provide objective signals for code validation and repair.
• Human Feedback: Interactive requirement clarification, code review, and preference-based RLHF/DPO pipelines align agent outputs with human intent.
• Self-Refinement: Agents employ self-critique, multi-agent collaborative feedback, and memory-based reflection to autonomously improve outputs.

  Figure 6: Feedback loops for coding agents, including internal self-refinement and external feedback from compiler, execution, and human sources.

Taxonomy of Vibe Coding Development Models

The paper introduces a three-dimensional classification of Vibe Coding development models:

1. Unconstrained Automation Model (UAM): AI-dominated, minimal human review, suitable for rapid prototyping but high technical debt risk.
2. Iterative Conversational Collaboration Model (ICCM): Human-in-the-loop review and understanding, analogous to pair programming, balancing speed and quality.
3. Planning-Driven Model (PDM): Upfront architectural planning, AI implements under strict human-defined constraints, suitable for complex systems.
4. Test-Driven Model (TDM): Test-first development, AI generates code to pass predefined tests, providing objective quality assurance.
5. Context-Enhanced Model (CEM): Horizontal enhancement via retrieval-augmented generation and context engineering, improving codebase alignment and maintainability.

These models are composable, allowing practitioners to tailor workflows to project risk, speed, and governance requirements.

Figure 7: Comparison of Vibe Coding development models with traditional software engineering models.

Open Challenges and Future Directions

Process Reengineering

Vibe Coding compresses the software development lifecycle into continuous micro-iterations, blurring traditional SDLC phases. Developers' roles shift toward intent articulation, system-level debugging, context curation, and architectural oversight. Project management and collaboration require new metrics, review processes, and accountability frameworks.

Code Reliability and Security

The abstraction of code authoring introduces risks of subtle bugs and vulnerabilities. Manual review is inadequate; integrated, real-time SAST/DAST, sandboxed dynamic analysis, and AI-driven threat modeling must be embedded in the development loop. The human remains the final arbiter, but automation must provide actionable, immediate intelligence.

Scalable Oversight

As agents become more autonomous, scalable oversight architectures are required. Weak-to-strong supervision, multi-agent debate, and continuous monitoring are necessary to prevent cascading errors, dependency proliferation, and alignment failures. Formal verification, RL-based watchdogs, and automated provenance tracking are critical research directions.

Human Factors

Developers transition from code authors to context engineers and supervisors. Prompt engineering, task decomposition, quality supervision, and agent governance become core competencies. Team collaboration and trust calibration must adapt to the scale and pace of AI-generated code, with implications for education and organizational structure.

Conclusion

This survey establishes Vibe Coding as a principled discipline, formalizing its triadic system, synthesizing agentic and infrastructural advances, and providing a unifying taxonomy of development models. The central insight is that context engineering, feedback integration, and human-in-the-loop governance are as critical as model quality for reliable, scalable, and maintainable AI-augmented software engineering. The paper articulates actionable guidance and a research agenda spanning technical, security, and human factors, positioning Vibe Coding as a foundational paradigm for the future of software development.