I’ve been bewildered by the wide range of experiences that software developers observe when working with AI. This has led many of us to outright reject AI for coding.
There are reasons for this variance.
Skeptics want to lightly dip their foot in the water to see what it’s about. This manifests as using the cheapest or free models and “just send it” styles of programming. Then they lament that what comes out is not what they would have built nor is it what they wanted.
Meanwhile, professional developers adopting agentic AI for coding operate differently. They see spending $100+/mo on AI tools as a great bargain. They spend hours planning and refining and decomposing a problem before they send a single request to their fancy AIs. Code carefully matching their request is often the outcome.
Is it a surprise they are getting different results? I’ve seen both styles and it’s no surprise to me.
These two experiences often get lumped under the same term: Vibe Coding.
They are not the same.
Today I ran across an essay by Addy Osmani, “Agentic Engineering.”
Addy really nails this difference and spells out a term that I’m keen to adopt: Agentic Engineering.
Vibe coding means going with the vibes and not reviewing the code. That’s the defining characteristic. You prompt, you accept, you run it, you see if it works. If it doesn’t, you paste the error back and try again. You keep prompting. The human is a prompt DJ, not an engineer.
…
Here’s the thing: a lot of experienced engineers are now getting massive productivity gains from AI - 2x, 5x, sometimes more - while maintaining code quality. But the way they work looks nothing like vibe coding. They’re writing specs before prompting. They’re reviewing every diff. They’re running test suites. They’re treating the AI like a fast but unreliable junior developer who needs constant oversight. I’ve personally liked “AI-assisted engineering” and have talked about how this describes that end of the spectrum where the human remains in the loop.
…
It draws a clean line. Vibe coding = YOLO. Agentic engineering = AI does the implementation, human owns the architecture, quality, and correctness. The terminology itself enforces the distinction.
Give the full article a read.
That last quote is important. Many devs dread AI because they don’t want to be a full time reviewer for marginal code. But if “AI does the implementation, human owns the architecture, quality, and correctness”, that sounds like senior developers’ job descriptions to me.
It’s something to think about.