Vibe Coding vs Agentic Engineering: What Actually Changes When You Ship
For a while I used the two terms interchangeably. Most people still do. They sound like the same activity with different branding: a human and an AI making software together, fast.
They are not the same activity. They produce different code, in different ways, with very different consequences. We run one of them in production now, on real client work. It took us a while to understand why the other one keeps ending in tears.
Here is the difference, in three parts.
Part I: One you get attached to. The other you throw away.
Vibe coding is when you stay up until 2am watching something work and having no idea why. It feels like magic. And because it feels like magic, you get attached to it. So when it breaks, you try to fix it. You patch the patch. You do, then you wait, then you fix, because throwing it away feels like losing something you made.
Agentic engineering takes that attachment out, on purpose.
The current wave of agentic development is built around two things: loops and a harness. The loop is the agent working through tasks on its own, again and again, while you do something else. The harness is everything wrapped around that loop to keep it honest: code styling rules, frontend conventions, security checks, a PR reviewer agent, and a human on the final review.
Once you trust the harness, you stop caring about any single run. I wake up to fifteen open pull requests. I open a few. I see the agent reached for something we would never ship, or solved a problem in a way that ignores our patterns. So I bin the whole thing and regenerate. No grief. Even if half the overnight tokens are gone with it.
In vibe coding you fix because you are attached. In agentic engineering, trash and regenerate is the default. (At least while tokens are cheap. That math may change.)
Part II: One is for thinking. The other is for building.
This is the part people miss, and it is in fact the part that makes me defend vibe coding rather than dismiss it.
Vibe coding is genuinely useful. Not for shipping, but for reasoning. You explore the problem by building a rough version of it. You find out what you actually want by watching a bad version of it run. It is a pencil and paper for a builder. The thinking happens through the making.
Agentic engineering is closer to traditional software development than it is to that. You do the thinking first, away from the keyboard. You decide what to build, how it should look, what the architecture is, the exact user journeys, the screens, sometimes down to the buttons. Then you hand that to the agents.
The workflow we use now looks like this. I have a conversation or a meeting with a colleague about a project. I dictate to Claude to pull the transcript, summarise it, and write a brief. Then I spend time with that brief. This is the important step, and it can take minutes, hours, or occasionally days. When the brief is right, I hand it to a Claude Code skill we built. It reads the brief, structures its own docs, breaks the work into tasks, and loops over them while I am asleep or doing something else.
The agents are fast. The thinking in front of them is still slow, and it is still mine. The speed at the end is only ever as good as the clarity at the start.
Part III: One does not scale. The other is how we ship now.
Vibe coding is fine for a prototype. It is fine for a throwaway internal tool nobody really depends on. It is not fine for a business core system.
We have a client right now running their entire operation on a vibe-coded solution. I will be direct about it: it is not good. It is full of bugs the original dev shop cannot fix, because the code underneath is a mess. They fix one thing and break two others. There were no standard practices anywhere in it. No PR reviews. No end to end tests. No security audit. It works until it doesn't, and then nobody can tell you why.
Agentic engineering is the opposite, because it does not throw away the last twenty years of software engineering. It uses AI inside that discipline, not instead of it. You build an architecture of agents, skills, and plugins that work together. The point is not to bypass the humans. The point is to make a small number of senior people far more capable than they were.
The results are not subtle. Spec to production now takes days, where it used to take weeks or months. That cuts the cost of building. And when it is done right, with the harness in place, it raises the quality bar rather than lowering it, because the checks never get tired and never skip a step under deadline.
So we are building more than we ever have. At a standard we are comfortable putting our name on.
TL;DR version
Vibe coding is a way to think with code. Agentic engineering is a way to ship with it. One is a sketchpad. The other is a workshop with rules.
The mistake is handing someone the sketch and calling it the building.
