Methodology

Vibe Coding vs Agentic Engineering: Same Tools, Opposite Disciplines

Both terms come from Andrej Karpathy, and both describe building software with AI agents. The difference is everything around the agent: no process versus real engineering process. Here is where each one fits, where each one breaks, and how you move from one to the other.

By Kylian Migot · Updated July 2026 · 7 min read

Quick answer

Vibe coding and agentic engineering use the same AI coding agents with opposite discipline. Vibe coding means accepting agent output without careful review, right for prototypes and throwaway projects. Agentic engineering wraps the same agents in specs, approval gates, isolation, and PR review, right for anything you have to maintain. Andrej Karpathy coined both terms and frames them as floor and ceiling, not rivals.
Vibe coding
Prompt, accept the diffs, ship the prototype (Karpathy, Feb 2025)
Agentic engineering
Same agents under specs, gates, isolation, and review (Karpathy, early 2026)
The dividing question
Will anyone have to maintain this code?
Karpathy's framing
Vibe coding raises the floor, agentic engineering raises the ceiling (Ascent, April 2026)
01

Two Terms, One Coiner

Vibe coding is Andrej Karpathy's February 2025 coinage for a way of building with AI where you give in to the vibes: prompt the model, accept the diffs without really reading them, and keep going, an approach he explicitly framed as fine for throwaway projects. The term escaped its origin, became shorthand for unreviewed AI-generated code in general, and was named Collins Dictionary's Word of the Year for 2025.

Agentic engineering is the successor term Karpathy proposed in early 2026, declaring his own coinage passé. It names the same activity, directing AI agents that write and run code, done with real engineering controls: a written spec before code, a human approval gate, isolated branches per task, verification against acceptance criteria, and PR review. The full definition and its six practices are covered in what is agentic engineering.

Because one person named both eras, this is not a turf war between camps. At Sequoia's Ascent event in April 2026, Karpathy framed the relationship himself: vibe coding raises the floor, agentic engineering raises the ceiling. The rest of this page is about the practical question that framing leaves open: which one should you be doing, on which work, today?

02

The Comparison, Dimension by Dimension

The table below is the whole argument in compressed form. Note the last two rows: both approaches have a real failure mode, and both have work they are genuinely right for.

DimensionVibe codingAgentic engineering
IntentPrototype: prove the idea fastProduction: software you must maintain
Code reviewSkipped, diffs accepted unreadGated, every change is a reviewed PR
SpecsNone, the prompt is the planWritten and enforced before code starts
Failure modeAn unmaintainable app nobody understandsProcess overhead on trivial tasks
When it's rightWeekend project, demo, one-off scriptAnything maintained, shared, or shipped

The dividing question is not tool quality or model choice, both columns can run the exact same Claude Code session. It is: will anyone have to maintain this code? If no, review is ceremony and you should skip it. If yes, every unread diff is a small loan against the future, and the workflow that pays those loans on time is the one described in engineering with AI agents.

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

03

When Vibe Coding Is the Right Call

A comparison page written by a company that sells discipline should be honest about this: vibe coding is often correct, and pretending otherwise is how methodology content loses your trust. Skipping specs and review is a rational trade whenever the cost of a latent defect rounds to zero.

Throwaway prototypes

You are testing whether an idea is worth building at all. The artifact's job is to answer a question, then be deleted. Review effort here is waste by definition.

Demos and spikes

A demo for Friday's meeting, a spike to see if a library fits. The output informs a decision; it never ships. Vibes are the appropriate quality bar.

Personal tools with one user

A script that renames your files or scrapes your own data. When you are the only user and the only victim, the feedback loop is instant and the blast radius is you.

Learning a new stack

Letting an agent build fast and reading the result afterward is a legitimate way to explore an unfamiliar framework. The deliverable is your understanding, not the code.
04

The Graduation Path: Prototype to Production

The most useful question is not “which side are you on” but “how does work cross the line.” A vibecoded prototype that earned real users does not need to be apologized for, it needs to graduate. The path that works treats the prototype as a requirements document, not a foundation:

  1. 1

    Freeze the prototype as the reference

    Stop adding features to the unreviewed codebase. Its remaining job is to demonstrate the behavior you want, which is the one thing vibecoded software is verifiably good at.
  2. 2

    Spec the rewrite, module by module

    Turn each behavior into a written spec: scope, files, acceptance criteria, exclusions. The prototype makes this unusually easy because you can describe what already exists instead of imagining it. The format is covered in spec-driven AI development.
  3. 3

    Gate the agents

    From here, no agent codes without an approved spec, and every task runs on its own branch or worktree. The same agents that built the prototype now work under supervision, and the difference shows up in the diffs.
  4. 4

    Verify and review every change

    Agents run tests against the acceptance criteria before handing off; you review each PR against its spec. Once a module passes review, it is production code, and the prototype's version of it is dead weight you delete.

Teams that skip the graduation and keep prompting against the prototype usually rediscover why: unreviewed code on top of unreviewed code compounds, and the eventual forced rewrite costs more than the deliberate one would have.

05

Where AIDEN Sits

AIDEN is built for the right-hand column of the table and for the graduation path. It is a macOS desktop app that orchestrates your local Claude Code and Codex CLIs on a kanban board, and its defining choice is an enforced spec gate: AIDEN drafts a spec from your story and its codebase analysis, and no agent starts coding until you approve it. Each story then runs in its own git worktree and lands as a PR with the spec attached, which is the loop described in Claude Code orchestration, made non-optional.

That makes AIDEN a poor vibe coding tool on purpose. If you are prototyping, a raw agent in a terminal is the right amount of process. The moment the work has to be maintained, the gate stops being friction and starts being the point. It is free for one project, so the graduation path above is cheap to trial on one real module.

FAQ

What is the difference between vibe coding and agentic engineering?
They use the same AI coding agents with opposite discipline. Vibe coding, coined by Andrej Karpathy in February 2025, means prompting an agent and accepting its diffs without careful review, which is fine for throwaway projects. Agentic engineering, the successor term Karpathy proposed in early 2026, wraps the same agents in process: a written spec before code, a human approval gate, isolated branches, verification against acceptance criteria, and PR review. The first optimizes for speed on disposable code, the second for correctness on code you must maintain.
Is vibe coding bad?
No, it is bad only when misapplied. For a weekend prototype, a demo, a one-off script, or a tool with a single user, skipping review is a rational trade: the code is disposable, so the cost of latent defects is near zero. Karpathy's original framing said exactly that, it is fine for throwaway projects. The problems start when a vibecoded prototype quietly becomes a product with users, at which point the unreviewed code becomes a liability someone has to pay down.
Can I turn a vibecoded prototype into production software?
Yes, but treat it as a graduation, not a continuation. Keep the prototype as a working reference of the behavior you want, then write specs that describe that behavior explicitly, and rebuild or refactor module by module with agents working under a gate: approved spec, isolated branch, tests, PR review. What rarely works is continuing to prompt against the unreviewed codebase and hoping quality emerges, more unreviewed code on top of unreviewed code compounds the problem.
Which is faster, vibe coding or agentic engineering?
Vibe coding is faster to first demo, agentic engineering is faster to reliable software. Skipping specs and review removes real steps, so a vibecoded prototype can exist in hours. But on maintained code the time comes back with interest: debugging code nobody read, untangling regressions, rewriting modules that cannot be extended. Per feature shipped and still working a month later, the disciplined loop usually wins, which is Karpathy's floor-and-ceiling point from Sequoia's Ascent event in April 2026.
Did the same person really coin both terms?
Yes. Andrej Karpathy coined vibe coding in February 2025, the term grew into shorthand for unreviewed AI code and became Collins Dictionary's Word of the Year for 2025, and in early 2026 he declared it passé and proposed agentic engineering as the successor. That makes this comparison unusually clean: the coiner himself frames them as stages of the same practice, not competitors.

Keep reading

Keep the agents, add the discipline.

AIDEN runs Claude Code and Codex on a kanban board with an enforced spec gate, so production work never starts on vibes. Free for one project.

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local