Meaning & Concepts

Vibe Coding vs Agentic Engineering: same AI, different job

A year after he coined "vibe coding", Andrej Karpathy started using a new term: "agentic engineering". Here's what each phrase actually means, how the two ways of working differ day to day, and how to tell which one your project needs.

By VibeCodingMeaning.com·Updated November 2026·17 min read

Two builders side by side at a dark desk, one casually prompting an AI chat on a laptop, the other reviewing an AI coding agent's task plan and test results on a large monitor — Both use AI to write the code. The difference is how much planning, checking and responsibility comes with it.

Vibe coding vs agentic engineering, in one sentence: vibe coding means letting an AI write code while you mostly ignore the code itself and judge the result by whether it seems to work, while agentic engineering means directing AI coding agents to do most of the writing while you plan the work, set the quality bar, and review and test everything before it ships. Both lean heavily on AI. What changes is who is accountable for the code, and how carefully it gets checked.

The distinction matters because the two words are often used as if they mean the same thing. They don't. Vibe coding is a great way to build a weekend game or test an idea in an afternoon. Agentic engineering is how professional teams are learning to use the same AI tools on software that real customers, payments and personal data depend on. This guide explains both in plain English, with a comparison table, a side-by-side example, prompts you can copy and a simple way to choose.

Quick answer

Vibe coding (Karpathy, February 2025): describe what you want, accept what the AI writes, and "forget that the code even exists". Fast, fun and great for prototypes.
Agentic engineering (Karpathy, February 2026): you rarely write code by hand; you orchestrate AI agents that do, and act as the oversight, "without any compromise on the quality of the software".
The real difference is review and responsibility: plans, tests, version control and someone who understands what shipped.
Not a replacement: vibe coding is still the right tool for throwaway projects. Agentic engineering is what you graduate to when the stakes rise.

What vibe coding means

Computer scientist Andrej Karpathy coined "vibe coding" in a post on X on February 2, 2025. He described a style of building where you "fully give in to the vibes" and "forget that the code even exists": you tell an AI what you want in plain language, run the result, and if something breaks you paste the error back in and ask for a fix. You don't read the code line by line. You judge the app by how it behaves.

The phrase spread fast. Collins Dictionary named it its Word of the Year for 2025, and Merriam-Webster added it in its 2026 update, defining it as "the act or practice of using an artificial intelligence system to generate computer code". If you want the full story, our complete guide to what vibe coding is covers the origin, the tools and the risks, and the vibe coding meaning homepage has the short version.

Notice the gap between those two definitions. Karpathy's original meaning was narrow: a specific, relaxed way of working where you don't look at the code. The dictionary meaning is broad: almost any use of AI to generate code. That gap is exactly why a second term became useful.

What agentic engineering means

In early February 2026, roughly a year after the original post, Karpathy wrote on X that his current favorite term for the professional version of this work was "agentic engineering". As reported by The New Stack, he explained both words:

"Agentic" "because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight".
"Engineering" "to emphasize that there is an art & science and expertise to it".

He also described the goal: "to claim the leverage from the use of agents but without any compromise on the quality of the software." In other words, keep the speed of AI, drop the "forget the code exists" part.

agentic engineeringnoun

/eɪˈdʒɛn.tɪk ˌɛn.dʒɪˈnɪə.rɪŋ/ · related: coding agents, AI-assisted engineering, vibe coding

Building software by directing AI coding agents that write, run and revise code, while a human plans the work, sets quality standards and reviews the results.
In Simon Willison's usage: "developing software with the assistance of coding agents", producing code that has been reviewed and brought up to a production-ready standard.

Developer Simon Willison, one of the most widely read writers on AI-assisted programming, started a project called Agentic Engineering Patterns on February 23, 2026: a living guide of practices for getting good results from coding agents. His definitions are some of the clearest available:

A coding agent is an AI that can both write and execute code. In his words, "Agents run tools in a loop to achieve a goal."
Running the code is what makes the approach work: the agent can try something, see the result, and iterate toward software that demonstrably works.
The human's job is to give the agent the right tools, describe the problem in enough detail, and verify and iterate on what comes back.

Willison also argues that "vibe coding" should keep its original, narrow meaning, which he describes as "unreviewed, prototype-quality LLM-generated code", rather than becoming a label for all AI-assisted coding. He made a similar point back in March 2025 in a post titled "Not all AI-assisted programming is vibe coding".

Vibe coding vs agentic engineering: side by side

Here's how the two approaches compare on the things that matter most in practice.

	Vibe coding	Agentic engineering
Who writes the code	The AI	The AI (agents), most of the time
Does a human read the code?	Usually not; you judge the result by how it behaves	Yes, or at least reviews diffs, tests and key decisions
Starting point	A quick prompt or idea	A plan or written spec with clear success criteria
Typical tools	Browser app builders and AI chat	Coding agents and AI code editors, plus Git and tests
How "done" is decided	"It seems to work"	Tests pass, review is done, requirements are met
Testing	Clicking around the preview	Automated tests the agent runs, plus manual checks
Version control	Optional; often none	Standard; every change is a reviewable commit
Who it suits	Beginners, non-programmers, anyone prototyping	Developers and teams shipping software people rely on
Best for	Toys, demos, personal tools, testing ideas	Production apps, customer data, payments, long-lived code
Biggest risk	Hidden bugs and security holes nobody looked for	Slower start; needs skill to plan and review well

The short version: the AI's job is roughly the same in both. The human's job is very different.

Why a new term appeared

If vibe coding was such a hit, why did its inventor reach for another phrase a year later? Three reasons stand out from the sources.

1. The tools changed. In early 2025, most people "vibe coded" in a chat window or an app builder, accepting one response at a time. By 2026, coding agents could take a task, edit many files, run the code and tests, read the errors and try again, largely on their own. Karpathy's "99% of the time" line describes that new default: the human is no longer typing code, but someone still has to direct and check the work.

2. "Vibe coding" stretched too far. Once dictionaries defined it as almost any AI code generation, the word stopped telling you whether anyone had actually checked the code. A senior engineer carefully reviewing an agent's pull request and a beginner shipping an app they've never looked at were both "vibe coding". A second term restores the difference.

3. Quality and security. Unreviewed AI code has real costs. Veracode's 2025 GenAI Code Security Report tested more than 100 AI models on 80 coding tasks and found that the generated code introduced security flaws in 45% of cases. "Engineering" signals that someone is responsible for catching those problems before users do.

It's worth noting what professional developers said at the time. In Stack Overflow's 2025 Developer Survey (more than 49,000 responses), about 77% said vibe coding wasn't part of their professional work, even though 84% used or planned to use AI tools. A summary of the same survey by the RDEL newsletter noted that AI agents were "not mainstream" yet, with 52% of respondents not using them. Agentic engineering is, in part, a name for the careful, professional way of using agents that many developers were still working out.

The same app, built both ways

Say you want a simple booking app for a small yoga studio: a class timetable, a sign-up form and an admin page for the owner.

The vibe coding version

You open an app builder and type: "Build a booking app for a yoga studio with a weekly timetable, sign-up form and an admin page."
A working preview appears in a minute or two. You click around. The timetable looks great.
The sign-up button errors, so you paste the error in. The AI fixes it.
You ask for a darker color scheme and a logo spot. Done.
You publish. Elapsed time: an afternoon. You never opened the code.

For a demo to show the studio owner, that's a great result. For a live system holding customers' names, emails and payments, it leaves big questions open. Can a logged-out visitor open the admin page? Can one customer see another's bookings? Nobody knows, because nobody checked.

The agentic engineering version

Plan first. You ask the agent to write a short spec: pages, data, who can see what, and how you'll know each feature works.
Review the plan. You fix anything wrong ("only the owner can see the full customer list") before any code exists.
Build in small tasks. The agent implements one feature at a time, on a Git branch, writing tests alongside the code and running them.
Review each change. You (or a developer) read the diff and the test results, and ask the agent to explain anything unclear.
Check the risky parts by hand. Logged-out test, two-account test, payment flow, error messages.
Ship and keep a trail. Merge, deploy, and keep notes so the next change starts from a known state.

It takes longer, though often not by as much as you'd expect, because the agent does most of the typing and testing. The difference is that at the end, someone actually knows how the app works and what it does with people's data.

Feel the vibe coding side first

Our Vibe Arcade has nine free games and mini apps that all started as plain-English prompts. Try the Prompt → App demo and watch an idea turn into a working toy.

Open the Arcade

The agentic engineering loop

Developer at a desk watching an AI coding agent work through a task list, with a test results panel glowing green and a code diff on a second screen — Plan, delegate, test, review, repeat: the agent does the typing, the human keeps the standard.

Different people describe it in different ways, but most agentic engineering workflows share the same loop:

Specify. Write down what you want, what "done" means and what must never happen (for example, "never expose customer emails to other customers").
Plan with the agent. Ask it to propose a plan and list its assumptions before writing code.
Delegate small tasks. One feature or fix at a time is easier to review than a giant rewrite.
Let the agent run the code. Tests, builds and the app itself. This is what makes it "agentic" rather than chat.
Verify. Read the diff, look at test results and try the feature yourself.
Commit. Save each good state in version control so you can always go back.

Willison's guide includes practical patterns that fit this loop, such as "red/green" test-driven development (have the agent write a failing test first, then make it pass), "first run the tests" before changing anything, using Git with coding agents, and asking the agent for "linear walkthroughs" that explain code you didn't write. Here's a prompt that bakes several of those habits in:

agentic-task.txt

// Paste into a coding agent at the start of a task
"Before writing any code:
1. Read the project and run the existing tests. Tell me
   if anything already fails.
2. Restate this task in your own words and list your
   assumptions and open questions.
3. Propose a short plan, broken into small steps.
Wait for my OK. Then, for each step: write a failing test
first, make it pass, run the full test suite, and commit
with a clear message. Finish with a plain-English summary
of what changed, what you tested and anything risky."

And when the agent finishes, this one helps you actually understand what shipped:

walkthrough.txt

// Use after a task is done
"Give me a linear walkthrough of the changes you just made,
file by file, in plain English. For each change, explain why
it was needed, what could break, and how the tests cover it.
Flag anything touching logins, payments, personal data or
secret keys."

For more ready-made prompts, see vibe coding prompts that actually work.

When to vibe code, and when to engineer

Neither approach is "right" in general. The question is what's at stake if the code is wrong.

Your project	Good fit	Why
A weekend game, toy or art project	Vibe coding	If it breaks, nobody is hurt; speed and fun matter most
Testing whether an idea is worth building	Vibe coding	You'll probably throw the first version away
A personal tool only you use	Vibe coding, plus basic care	Low risk, but back up anything important
A demo to show investors or a client	Vibe coding, clearly labeled as a prototype	Fine for showing; don't let it quietly become the product
An internal tool your team depends on	Agentic engineering (light)	Bugs cost people time; keep version control and tests
Anything with logins or customer data	Agentic engineering	Access control mistakes are common and serious
Payments, health or financial information	Agentic engineering plus expert review	The cost of a mistake is high, and rules may apply
Code that will be maintained for years	Agentic engineering	Someone has to understand it later

A useful habit: start projects in vibe coding mode, and switch to engineering mode the moment the project gets real users, real data or real money. Our vibe coding security checklist is a good bridge between the two.

Is vibe coding dead?

Headlines in 2026 had fun with the idea that Karpathy "retired" vibe coding. That's not quite what happened. His February 2026 comments were about the professional default: for serious software, the expectation is now agents plus oversight, not unchecked output. Nothing in them says casual, low-stakes vibe coding stopped being useful.

A neat way to think about it comes from a May 2026 write-up on Travis Media: vibe coding "raises the floor" (more people can build software at all), while agentic engineering "raises the ceiling" (professionals can build more, and better, with agents). The same write-up quotes a line from a 2026 Karpathy talk that sums up the whole debate: "You can outsource your thinking, but you can't outsource your understanding."

So vibe coding isn't dead. It has a clearer job description. It's how millions of non-programmers get started, and how professionals sketch ideas quickly. Agentic engineering is what happens when those sketches need to become dependable software. If you're wondering where you fit, our guide to the vibe coder meaning breaks down the different kinds of people who call themselves vibe coders, and vibe coding vs traditional coding covers whether to learn programming at all.

How to move from vibe coder to agentic engineer

Self-taught builder at a laptop in a dim room studying a code diff with an AI assistant's explanation open beside it, notebook of checklists on the desk — The skills that matter most are planning, reading what changed, and testing it, not typing code from memory.

You don't need a computer science degree to start working more like an engineer. You need a handful of habits, most of which the AI can teach you as you go:

Learn basic Git. Ask your AI tool to set up version control and explain commits, branches and how to roll back. It's the single biggest safety net.
Write a one-page spec before big features. What it does, who can see what, and how you'll know it works.
Ask for tests, and run them. Even a few tests for the most important flows (sign-up, checkout, permissions) catch a lot.
Read the diff, not just the preview. You don't have to understand every line. Ask the agent to walk you through it and flag anything risky.
Keep tasks small. "Add password reset" is reviewable. "Rebuild the whole app" usually isn't.
Learn the security basics. Secret keys stay out of browser code, database rules are locked down, and you test logged out and with two accounts. Our guide on how to deploy a vibe-coded app walks through these checks.
Pick tools that support the loop. Coding agents and AI code editors make plans, tests and Git easy; see our roundup of the best vibe coding tools or browse the tools directory.

Want a quick gauge of where you are now? Take the two-minute Vibe Challenge. Need a project to practice on? Spin the Vibe Machine idea generator. And if your project has outgrown the "vibes" stage but you'd rather not do the engineering yourself, you can hire a vibe coder to build, test and hand it over properly.

Other terms you'll hear

The vocabulary is still settling, so here's a quick guide to the neighbors of these two terms:

AI-assisted coding / AI-assisted engineering: the broad umbrella for any programming where AI helps, from autocomplete to agents. Willison's point is that most of it isn't vibe coding.
Agentic coding: often used interchangeably with agentic engineering, usually focused on the act of working with coding agents rather than the whole discipline.
Coding agent: an AI that can edit files, run commands and tests, and iterate on its own toward a goal.
Spec-driven development: writing a clear specification first and having agents build against it. It's one common style of agentic engineering.
No-code: building with visual building blocks instead of generated code. We compare it in vibe coding vs no-code.

Frequently asked questions

What is the difference between vibe coding and agentic engineering?

Vibe coding means letting AI write code while you mostly ignore the code and judge the result by whether it seems to work. Agentic engineering means directing AI coding agents to write most of the code while you plan the work, set quality standards, and review and test the results before they ship.

Who coined the term agentic engineering?

Andrej Karpathy, who also coined "vibe coding", said in a post on X in early February 2026 that "agentic engineering" was his favorite term for the professional version of AI-driven development. The New Stack reported his explanation on February 10, 2026.

What did Karpathy mean by "agentic engineering"?

He said "agentic" reflects that you no longer write the code directly 99% of the time but orchestrate agents and act as oversight, and "engineering" emphasizes that there is an art, science and expertise to it. His stated goal was getting the leverage of agents without compromising software quality.

Is vibe coding dead now?

No. Agentic engineering describes the professional standard for serious software, but vibe coding is still a fast, useful way to build prototypes, games, demos and personal tools where a hidden bug wouldn't hurt anyone.

Is agentic engineering the same as AI-assisted coding?

It's a specific kind of AI-assisted coding. AI-assisted coding covers anything from autocomplete to chat. Agentic engineering focuses on coding agents that can write and run code in a loop, with a human planning and reviewing the work.

What is a coding agent?

A coding agent is an AI tool that can edit code, run commands and tests, read the results and keep iterating toward a goal. Simon Willison describes agents as software that runs tools in a loop to achieve a goal.

Do I need to know how to code for agentic engineering?

It helps a lot, because the human's main job is judging whether the agent's work is right. Beginners can start by learning Git, asking for tests, and having the agent explain every change in plain English, then build up from there.

Which is better for a beginner, vibe coding or agentic engineering?

Start with vibe coding to learn what's possible and build confidence on low-stakes projects. Add agentic engineering habits such as specs, tests, version control and code review as soon as a project gets real users, data or money.

Is agentic engineering safer than vibe coding?

Usually, because the code gets planned, tested and reviewed. Veracode's 2025 research found security flaws in 45% of AI-generated code samples it tested, so having a human check the output before it ships matters. It's still only as safe as the review actually is.

What tools are used for agentic engineering?

Mainly coding agents and AI code editors that can run code and tests, used together with Git and a test suite. The specific tools change quickly, so compare current options and check each tool's own documentation.

Keep reading

Browse every guide on the vibe coding blog.

Sources

The New Stack, "Vibe coding is passé. Karpathy has a new name for the future of software", Darryl K. Taft, February 10, 2026: thenewstack.io
Glide Blog, "What is agentic engineering?", February 6, 2026 (quotes Karpathy's X post): glideapps.com
Simon Willison, "Writing about Agentic Engineering Patterns", February 23, 2026: simonwillison.net
Simon Willison, Agentic Engineering Patterns: "What is agentic engineering?": simonwillison.net
Simon Willison, "Not all AI-assisted programming is vibe coding (but vibe coding rocks)", March 19, 2025: simonwillison.net
Travis Media, "Vibe Coding vs Agentic Engineering: Andrej Karpathy's Big Idea for AI Coding", May 1, 2026: travis.media
Stack Overflow, 2025 Developer Survey: survey.stackoverflow.co
RDEL #105, "Top findings from the 2025 Stack Overflow Developer Survey": rdel.substack.com
Veracode, 2025 GenAI Code Security Report press release, July 30, 2025: veracode.com

VibeCodingMeaning.com is an independent educational site and isn't affiliated with any person, tool or platform mentioned. Quotes are taken from the linked sources; tools and terminology in AI coding change quickly, so check current documentation. Some links may be affiliate links.