L02 · Before you start · 5 min
You lead, the agent builds: agentic engineering in plain words
In one sentence
Agentic engineering is a structured way to build real software where you set the goal, agree on a written spec and plan, and direct a coding agent to build one small step at a time while you inspect and approve every result.
Think of it like this
Imagine directing a film: you don't hold every camera or sew every costume yourself, yet every scene exists because you approved the script, the storyboard, and the final cut. Without your direction, the crew would just shoot random footage.
How it works
In February 2026, AI researcher Andrej Karpathy proposed the term Definitionagentic engineeringBuilding software by directing AI agents and checking their work: you lead, the agent builds.Open in the glossary → to distinguish serious software building from “prompt and pray” (often called vibe coding). A Definitionquick experimentA throwaway try-out without a spec. Fine for testing an idea, never for something you'll publish.Open in the glossary → where you accept whatever the AI outputs is fine for a throwaway test, but real apps need a clear structure.
Instead of asking your Definitioncoding agentAn AI that works in your project folder: it reads files, writes code and runs commands, with your permission.Open in the glossary → to build an entire app in one giant leap, you follow Definitionspec-driven development (SDD)Agree on what to build (the spec) and a plan, then build one small, checked step at a time.Open in the glossary →: Idea → Spec → Plan → Build → Check → Publish. First you agree on what to build and why in a plain-language Definitionspec`SPEC.md`: what you're building, who it's for, what's in and out, and how you'll know it works.Open in the glossary → (SPEC.md), then you split that spec into a step-by-step DefinitionplanThe spec split into small tasks, each finished and checkable on its own, in `tasks/plan.md`.Open in the glossary → (tasks/plan.md).
During the Build and Check stages, the agent works on one small task at a time, saves a DefinitioncommitA checkpoint: a snapshot of what changed, with a short note. You can return to any commit.Open in the glossary → after each working step, and runs automated checks (such as a DefinitiontestA small program that checks your app does what it should, so a fix can't silently break later.Open in the glossary →) to prove nothing broke. You try the feature in your browser or in a private DefinitionPreviewA private test link Vercel makes for each change, so you can try it before it goes live.Open in the glossary → link before moving on.
You stay in charge because you own the decisions that matter: who the app is for, what stays out, how it should feel, and when it is ready to DefinitiondeployPutting a version of your app online.Open in the glossary → to DefinitionProductionThe live version of your app, the one everyone sees. It comes from `main`.Open in the glossary →. The agent handles the syntax and commands; you provide the judgment.
What it looks like in your project
At the start of every session, your agent reads SPEC.md and tasks/todo.md and shows you where you are on the path: ① Setup → ② New project → ③ Spec → ④ Plan → ⑤ Build & preview → ⑥ Publish.
When you ask for a new feature, it won't jump straight into writing code. Instead, it will interview you one question at a time to update the spec and plan first, waiting for your OK at each stage.
Copy for your agent
Read SPEC.md and tasks/todo.md (if they exist), show me where we are in the six-step path, and propose the single next step we should take together.
Quick check · 1 question