The method
You lead, the agent builds
Agentic engineering is building software by directing AI agents and checking their work. We do it the spec-driven way: agree on what to build, plan small steps, build one checked step at a time, and publish only when you say so.
Who does what
YouYou leadDecide what to build, approve each step, check the result.
Coding agentThe agent buildsReads files, writes code, runs commands. Asks first.
Your projectReal code, yoursFiles and checkpoints in your folder, then on your GitHub.
The loop, for every feature
- IdeaWhat and who for
- SpecAgreed in SPEC.md
- PlanSmall, checkable tasks
- BuildOne step at a time
- CheckTests and a Preview
- PublishOnly when you say so
The Path, for your first app
Why not “prompt and pray”?
In February 2025 Andrej Karpathy coined “vibe coding” for giving in to the AI and not reading the code. It became shorthand for apps that break and leak keys. In February 2026 he proposed “agentic engineering” for the professional practice: people orchestrate agents and oversee their work. We use that vocabulary, made beginner-friendly.
| Aspect | Prompt and pray | Agentic engineering |
|---|---|---|
| Who decides what to build | Whatever the AI produces | You, in SPEC.md |
| Size of each step | Everything at once | Small tasks, a checkpoint after each |
| Checking | “Looks fine” | Tests, a Preview, the quality bar |
| Secrets | Pasted wherever it works | Only in Vercel and .env.local |
| Going live | Whenever | Only with your OK, after review |
| When it breaks | Random fixes | Reproduce, find, fix, prove |
| What you learn | Very little | The ideas, just in time |
Sources
Was this page helpful?