# Build With a Coding Agent

**Track:** Foundations of AI-Native Development — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (coding agents)
**Level:** Beginner
**Prerequisites:** Git as Memory & Reversibility

**In one line:** Replace prompt-and-hope with a disciplined loop: specify → plan → build → run → inspect → verify → revise.

## Theory, aesthetics & inspiration

When Andrej Karpathy named "vibe coding" in early 2025 — conversational software, the model's output accepted without reading every line — he was describing a mode, not a method. The method is a controlled loop: specify → plan → build → run → inspect → verify → revise, with the agent as an operator inside it rather than an oracle above it. Give it a bounded objective, the relevant repository context, explicit constraints, and a definition of done; ask it to inspect before editing and to run the result after. The tooling has converged on this shape — by 2026, Next.js was shipping agent-oriented project context, diagnostics, and browser visibility in its 16.2/16.3 releases, on the observation that agents get more reliable when they can observe the real environment instead of inferring it from pasted snippets. The human role moves upward: choose the architecture, constrain the scope, judge the evidence. The operative skill is not prompting; it is directing a tool-using collaborator.

**Builder question:** What evidence should the coding agent produce before you accept that a feature is actually finished?