Vibe Coding vs Vibe Engineering vs Agentic Engineering: The 2026 Glossary

Vibe coding, vibe engineering, and agentic engineering are three ways of building software with AI, and they differ by how much you review and own the result. Vibe coding (Andrej Karpathy, February 2025) means prompting an AI and shipping what it gives you without deep review. Vibe engineering (Simon Willison, October 2025) means using AI to go fast while staying fully accountable for the code. Agentic engineering (Karpathy, April 2026) means acting as the architect who writes the spec and directs AI agents that produce almost all of the implementation under your oversight.

設定をやめて、構築を始めよう。

AIオーケストレーション付きSaaSビルダーテンプレート。

Why this matters to you

The label is not the point. The accountability is. Each step up the ladder asks you to own more of the security, correctness, and architecture of what you ship. Pick the wrong one for the job and you can leak data or break payments. Pick the right one and you move fast without shipping something dangerous.

Two people coined all three terms in about 14 months: Andrej Karpathy (who helped start OpenAI and led AI at Tesla) and Simon Willison (creator of the Django web framework and the Datasette data tool). That is unusual, and it tells you the vocabulary is still forming as AI agents get more capable.

The one-screen comparison

Question	Vibe Coding	Vibe Engineering	Agentic Engineering
Coined by	Andrej Karpathy	Simon Willison	Andrej Karpathy
Date coined	Feb 2, 2025	Oct 7, 2025	Apr 30, 2026
Core definition	Prompt the AI, accept the output, ship without close review	Use AI to move fast while reviewing and owning every change	Act as architect: write the spec, direct AI agents, own correctness and security
Who it is for	Anyone, including non-coders	Experienced engineers	Experienced engineers and teams
Human role	Describe what you want	Reviewer and decision-maker	Architect and orchestrator
AI role	Writes most of the code	Writes code you check line by line	Plans, builds, and tests ~99% of the implementation
When to use it	Prototypes, demos, personal tools	Production apps with real users	Regulated, payments, auth, enterprise

Vibe coding: the floor

Karpathy coined "vibe coding" in an X post on February 2, 2025 that reportedly passed 4.5 million views. He described giving in to the vibes, accepting what the AI writes, and barely reading the diffs. Collins English Dictionary named it Word of the Year for 2025.

The honest framing: vibe coding raises the floor. It lets people who could never build software before make a working thing. That is real value. Willison later described the style as "fast, loose, and irresponsible," which is fine for a throwaway prototype and risky for anything real.

Plain example: you ask an AI to "build me a tip calculator," it does, you ship it to friends. Nobody gets hurt if it has a bug.

Vibe engineering: the responsible middle

Simon Willison coined "vibe engineering" on his site simonwillison.net on October 7, 2025. He drew an explicit contrast with vibe coding. Vibe engineering is for "seasoned professionals" who use AI to go faster while "staying proudly and confidently accountable for the software they produce."

This is the key idea most competitor posts skip entirely. Vibe engineering is not a rebrand of vibe coding. It exists because by late 2025 the AI was fast enough to tempt even good engineers into skipping review. Willison gave the skip a name to resist, not a name to celebrate. You read every change. You own the result. The AI is a fast typist, not the person on the hook.

In May 2026, Willison added an honest caveat: as coding agents get more reliable, even experienced engineers stop reading every single line, which blurs the line between vibe engineering and agentic engineering. The two are converging. That makes this less a fixed glossary and more a snapshot of a moving target.

Agentic engineering: the ceiling

On April 30, 2026 at the Sequoia AI Ascent event, Karpathy introduced "agentic engineering" and called vibe coding "passé." His framing: vibe coding raises the floor, agentic engineering raises the ceiling.

The model is human as architect-orchestrator. You write the spec. AI agents plan, build, and test. Quality gates enforce standards. The AI may produce around 99% of the implementation, but you own the outcome. Karpathy was blunt about the line you cannot cross: "You are not allowed to introduce vulnerabilities because of vibe coding. You are still responsible for your software, just as before."

This pairs with patterns you may already use, like writing a clear CLAUDE.md instructions file, running Claude Code subagents in parallel, and wiring up MCP servers so the agents can reach your tools.

Why the stakes are real

The terminology debate matters because AI-written code fails in specific, measurable ways.

AI co-authored code shows 2.74x higher security vulnerability rates than human-written code (CodeRabbit, reported December 2025).
CVEs (publicly tracked security flaws) attributed to AI-generated code rose from 6 to 35 in two months, January to March 2026 (reported).
Only 8.25% of AI outputs were both functionally correct and secure in one study (Veracode, reported).

Those numbers are the reason "who is accountable" is not a word game. The more code the AI writes, the more deliberate your review and your guardrails need to be.

How to pick your approach

A simple self-test by use case:

Prototype, hackathon, or personal tool: vibe coding is fine. Ship it.
Production app with real users: vibe engineering is the minimum. Review every change, own it.
Regulated, enterprise, payments, or authentication: agentic engineering. Use specs, agents, and automated security checks.

This is where a build system helps. Build This Now is the $29 Code Kit, a harness for Claude Code that ships a production skeleton (auth, Stripe payments, PostgreSQL with row-level security on every table). It wires agents to plan, build, and test against a spec you set, with quality gates and /security and /pentest commands that target exactly the CVE and OWASP failures vibe coding tends to create. That is agentic engineering in practice, not theory.

FAQ

What is vibe engineering?

Vibe engineering, coined by Simon Willison on October 7, 2025, means using AI tools to speed up development while staying fully accountable for the code you produce. It is the disciplined counterpart to vibe coding's accept-everything approach: you still review and own every change.

What is the difference between vibe coding and agentic engineering?

Vibe coding (Karpathy, February 2025) means prompting AI and shipping code without deep review. Agentic engineering (Karpathy, April 2026) means acting as architect-orchestrator: writing specs, coordinating AI agents, and owning correctness and security, with AI producing roughly 99% of the implementation under structured human oversight.

Is vibe coding safe for production apps?

No. AI co-authored code shows 2.74x higher security vulnerability rates than human-written code (CodeRabbit, reported December 2025), and Karpathy himself said "you are not allowed to introduce vulnerabilities because of vibe coding." Production apps need at minimum vibe engineering discipline, and anything handling payments or user data warrants full agentic engineering practices.

Who coined vibe coding and when?

Andrej Karpathy coined "vibe coding" on February 2, 2025 in an X post that reportedly reached 4.5 million views. Collins English Dictionary named it Word of the Year for 2025 in November of that year.

Vibe Coding vs Vibe Engineering vs Agentic Engineering: The 2026 Glossary

On this page