Vibe Coding vs Agentic Coding: The Real Difference
Vibe coding is about how much you look at: you judge the result and skip the code. Agentic coding is about how much the AI does by itself: it reads files, changes them and runs things in a row, while a person still answers for the quality.
Why are these not opposites?
They answer different questions, so you can do both at once. One is about you and the other is about the tool. Claude Code, for example, describes itself as "an agentic coding tool that reads your codebase, edits files, runs commands". If you use a tool like that and never look at the code, you are vibe coding with an agent.
The word got a sharper meaning in February 2026. A year after he coined "vibe coding", Andrej Karpathy wrote that many people had looked for a name to set serious work apart, and that "my current favorite is agentic engineering". You will see "agentic coding" and "agentic engineering" used for nearly the same thing. Addy Osmani sums up the idea in one line: "AI does the implementation, human owns the architecture, quality, and correctness."
| Vibe coding | Agentic coding | |
|---|---|---|
| Who reads it? | Nobody | A person, or a process that stands in |
| What counts as done? | It seems to play right | It meets what was written down first |
| Good for | Fun, tryouts, small games | Things people rely on |
Do I have to pick one for the whole game?
No, and it is better not to. Take an invented game where you collect creatures from tide pools and keep them in a book. Sort its parts by one question: how quickly would I notice if this were wrong, and how much would it hurt?
Keep it loose. The colour of the water, the way crabs wiggle, the title screen. If one of these is wrong you see it at once.
Be careful. The save file, the book that counts your 40 creatures, the shop prices. If the save file drops one creature in fifty, you may not notice for weeks, and then a player has lost something they cared about.
What does careful mean if I cannot read code?
You cannot review the code, but you can review the behaviour. For the careful pile, four habits go a long way:
- Say what it must do first. "A creature in the book stays there after quitting the game."
- Have it drawn before it is built. A diagram of boxes and arrows is something you can read, even when the code is not.
- Ask for reasons you can look at. The AI points to what it read and tried, or says it does not know.
- Use a second chat. It has no reason to defend a change it did not make.
Osmani's version has a human reviewing every change, and you cannot fully do that without reading code. So be honest about the limit. Once real money, accounts or other people's data are involved, get a programmer to look at it.
Paste this to your AI
Look through my game and sort every part into two piles.
LOOSE: if this is wrong I will see it at once and it costs nothing.
CAREFUL: if this is wrong I might not notice, and a player could lose something.
For each CAREFUL part, name the place in the project and tell me what the game does there today. Only sort what you have really opened. Anything you have not read goes under NOT READ, and you do not guess about it.
Then take the riskiest CAREFUL part. Before you suggest any change, chart it in Mermaid: the steps it passes through and what triggers each one, based on what you read. Walk every path and mark any that loses something. Tell me what could go wrong, what you saw that makes you think so, and one thing I can do in the game to see it happen. Any fix gets built from that diagram.
Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP
How to check it worked
- Look at NOT READ. Ask the AI to read those parts, or ask why it could not.
- Read the diagram. Can you follow a creature from the pool to the book and back after a restart?
- Quit and come back. Play, save, close the game, open it again. Is everything from the careful pile still there?
- Ask a fresh chat to sort the same game. Where the two lists disagree, you have found something worth a look.
For a full way of working, read the simple workflow.