💭 in your head · nothing to type
Run it in your head — the last task where a single prompt turned out to be a blunt tool where you needed a crew.
Don't picture it as "a prompt that failed" — picture it as a heist that one person tries to pull off alone. Picking the lock, killing the cameras, driving the getaway car, and counting the loot, all at once. Something always drops. A professional crew splits the roles — every specialist covers their lane.
Replay it: how many "roles" did you end up juggling yourself — or forcing one prompt to juggle? Which link was the most fragile?
💭 in your head · nothing to type
Andrej Karpathy (2026) drew a sharp line: vibe coding raises the floor, agentic engineering raises the ceiling.
Vibe coding = drops the entry barrier to zero. Any non-coder can build an MVP through ChatGPT in an evening. This is about the prototype.
**Agent engineering** = raises the ceiling of the profession. The discipline of assembling one reliable agent from roles, tools, and checks. This is about production.
You can outsource execution but not comprehension.
Don't picture this as programming — picture it as blacksmithing. The vibe coder grabs the hammer and hits the hot iron: one swing, see what comes out, swing again. The agent engineer first designs the tooling — the die, the press, the quench. Then that tooling does the forging, run after run, with a predictable result.
Hold the image: what in your task is a single hammer swing? And what deserves its own tooling?
Signs you've hit the ceiling of vibe coding:
| Symptom | What it means |
|---|---|
| Prompt works 7 out of 10 times | You need verification and retry |
| You need data from 3+ sources | You need a pipeline, not a prompt |
| Task repeats every day/hour | You need trigger • scheduling |
| Errors cost money/trust | You need observability and human-in-loop |
| You want someone else to run it | You need a spec and UI |
If even one of these applies — welcome to agent engineering.
Three eras of programming (Karpathy: Software 2.0 — a 2017 essay, Software 3.0 — a 2025 talk):
Software 1.0 → humans write code
Software 2.0 → humans write data, ML models "train"
Software 3.0 → humans write prompts, neural nets "execute"