Vibe Coding

Vibe coding is an informal AI-assisted development style where a person describes what they want in natural language, accepts or iterates on the generated code, and relies mostly on quick manual checks instead of structured specification, tests, evals, and architecture review [src-055, src-094].

Key points

• Karpathy popularised the term for a looser style of AI-assisted programming, where the developer follows the model's output and fixes issues by feeding errors back into the prompt [src-055, src-094].
• ^[src-094] treats vibe coding and Agentic Engineering as endpoints on a spectrum, not mutually exclusive tool categories.
• The production risk is weak verification: if the only question is whether the output "seems to work", the workflow remains vibe coding even if it uses advanced models or long prompts ^[src-094].
• Vibe coding can be useful for exploration, learning, prototypes, or low-stakes experiments, but it creates maintenance, security, and token-cost debt when used for serious systems without a harness ^[src-094].
• The practical upgrade path is to add structure: explicit specs, architectural context, tests, evals, guardrails, observability, and human judgement ^[src-094].

Related entities

• Andrej Karpathy
• Addy Osmani
• Shubham Saboo
• Sokratis Kartakis

Related concepts

• Agentic Engineering
• Software 3.0
• Context Engineering
• Harness Engineering
• Software Factory Model
• Continuous Agent Evaluation
• Claude Code Token Economics

Source references

• ^[src-055] Sequoia Capital – "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (2026-04-29)
• ^[src-094] Addy Osmani, Shubham Saboo, Sokratis Kartakis – "The New SDLC With Vibe Coding" (2026-05)