Vibe Coding vs. Natural Language Development (marmelab.com)

Agentic coding tools like Claude Code, GitHub Copilot Agent and Cursor have crossed a practical threshold: they now navigate large codebases, follow conventions and translate business requirements into working modules far faster than traditional autocomplete. The author contrasts a 30‑minute hand‑coded user registration using bcrypt and JWT with a 10‑minute AI‑generated authentication module (registration, login, password reset, email verification, rate limiting) to argue for a new specialist: the Natural Language Developer, who specifies requirements in plain English, reviews AI output, and iterates on behavior rather than writing syntax. This shift is significant because it reframes developer skills and workflows. Critical competencies will include precise specification writing (RFC‑grade validation rules, security checklists), behavior‑driven testing, agent orchestration, and living spec/versioning systems (Git‑style diffs for intent). Key risks remain: agents exhibit nontrivial failure modes (~5% logical errors), prompt‑injection and security vulnerabilities, and brittle generalization across edge cases. Practical practices recommended include incremental task decomposition, human‑owned BDD tests, redundancy or “quorum” strategies for AI outputs, and stronger architectural oversight. The ecosystem will bifurcate into spec‑first natural language developers, code‑centric engineers who debug and harden systems, and foundational builders who create the runtime and agent infrastructure—an ecosystem, not a replacement, with new higher‑level problems supplanting low‑level concerns.