Karpathy coined both terms a year apart. One builds $400M startups. The other lost Amazon 6.3 million orders. The difference is about to define which developers thrive.
Collins Dictionary named "vibe coding" its Word of the Year for 2025. Three months later, the man who coined the term declared it "passé" and replaced it with "agentic engineering." Between those two events, a Swedish startup called Lovable hit $400M ARR selling vibe-coded apps, Amazon lost 6.3 million orders to an AI-generated code deployment, and CodeRabbit found that AI-written code produces 1.7x more major bugs than human code.
Same technology. Wildly different outcomes. The difference isn't the AI. It's how you use it. And that difference is about to define which developers thrive and which become obsolete.
Both "vibe coding" and "agentic engineering" come from Andrej Karpathy -- co-founder of OpenAI, former AI director at Tesla, one of the most respected voices in machine learning.
February 2, 2025: Karpathy tweets:
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists... I 'Accept All' always, I don't read the diffs anymore."
February 8, 2026: Karpathy declares vibe coding "passé" and introduces "agentic engineering":
"'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."
One year. Same person. A complete philosophical reversal -- from "forget the code exists" to "there is an art and science and expertise to it." That arc tells you everything about where AI coding is headed.
Let's get the numbers on the table. They tell two stories simultaneously.
| Metric | Data | Source |
|---|---|---|
| YC W25 startups with 95%+ AI code | 25% | TechCrunch |
| Developers using AI tools daily | 51% | Stack Overflow 2025 |
| Code generated by Copilot (avg) | 46% | GitHub statistics |
| Copilot speed improvement (certain tasks) | 55% faster | GitHub research |
| Lovable time to $100M ARR | 8 months | TechCrunch |
| Metric | Data | Source |
|---|---|---|
| AI code with OWASP Top 10 vulns | 45% | Veracode |
| More vulnerabilities in AI code vs human | 2.74x | Apiiro |
| More major bugs in AI code | 1.7x | CodeRabbit |
| More privilege escalation paths | 322% | Apiiro |
| Experienced devs slower with AI tools | 19% | METR |
| Developers who trust AI tools | 29% | Stack Overflow 2025 |
Both stories are true. AI makes you faster at producing code and worse at producing correct code. The question is whether you treat that tradeoff as acceptable.
Simon Willison (Django co-creator) drew the critical distinction that most articles miss: "Vibe coding is NOT the same thing as writing code with the help of LLMs."
Vibe coding is a specific practice: you describe what you want in natural language, the AI generates code, and you accept it without review. You don't read the diffs. You don't understand the implementation. You trust the vibes.
This is different from using Copilot for autocomplete. It's different from asking Claude to explain a function. It's different from using an AI agent to implement a feature you've spec'd out and will review. Vibe coding means abdicating understanding.
Karpathy was explicit about this in his original tweet: "I 'Accept All' always, I don't read the diffs anymore." That's the defining characteristic. Not AI assistance. Blind acceptance.
Lovable's $400M ARR proves this. People want to build apps quickly. Lovable generates 100,000+ new projects daily. Most of those projects don't need production-grade code. They need something that works now.
The disaster list is long and growing. Here are the highlights from 2025-2026:
terraform destroy. 1.94 million database rows lost. 100,000+ students affected. 2.5 years of production data gone.The pattern is always the same: the code looks correct, passes basic tests, and hides a vulnerability that a human reviewer would have caught in minutes.
Apiiro's study of Fortune 50 enterprises found that AI-assisted developers produced 3-4x more code but also 10x more security issues. By June 2025, their monitored enterprises were generating 10,000+ new security findings per month -- a 10x spike in six months.
After just 5 iterative revisions, AI-generated code contained 37% more critical vulnerabilities than the initial generation. The code gets less secure the more you iterate on it with AI. That's terrifying.
Karpathy's shift from vibe coding to agentic engineering wasn't just a rebranding. It reflected a fundamental change in his workflow:
The key difference:
| Dimension | Vibe Coding | Agentic Engineering |
|---|---|---|
| Who writes the code | AI, unreviewed | AI, with human oversight |
| Developer's role | Describe the vibe | Architect, reviewer, QA |
| Code understanding | "Forget the code exists" | Deep understanding of what agents produce |
| Testing | Minimal or none | Rigorous, automated |
| Best for | Prototypes, throwaway code | Production systems, team codebases |
| Career risk | High (skill atrophy) | Low (skill amplification) |
In agentic engineering, you don't write the code. But you write the specs. You review the PRs. You design the architecture. You catch the bugs the AI introduces. You're a tech lead managing AI agents instead of junior developers.
Addy Osmani (Google) summarized the five principles:
That fifth point is the career-defining one. Vibe coders don't own the system. They can't, because they don't understand it. Agentic engineers own everything -- they just didn't type most of it.
The tooling split maps directly onto the vibe-vs-agentic divide.
Platforms like Lovable, Bolt, Replit Agent, and v0 are designed for vibe coding. You describe what you want, and you get a working app. No IDE. No terminal. No git. Just... a product.
These are excellent for what they're designed for. They're dangerous when used beyond their scope.
| Tool | SWE-bench Pro | Key Strength | Monthly Cost |
|---|---|---|---|
| OpenAI Codex CLI | 57.0% | Highest benchmark score, token-efficient | $20 + usage |
| Claude Code | 55.4% | Multi-agent, git integration, MCP | $20-200 |
| Cursor (Agent Mode) | 50.2% | Best autocomplete, VS Code familiarity | $16-20 |
| Windsurf | -- | Persistent context (Cascade), IDE plugins | $15-200 |
| Devin 2.0 | -- | Most autonomous, sandboxed environment | $20 + $2.25/ACU |
Source: Scale Labs SWE-Bench Pro Leaderboard
These tools don't generate apps from vibes. They integrate into professional workflows -- terminal, git, CI/CD, code review. They're designed for developers who understand what the AI is doing and can course-correct when it's wrong.
The distinction matters. Cursor's Agent Mode lets you review every change. Claude Code commits to git with proper messages. Codex CLI runs in a sandbox. These are guardrails that vibe coding platforms deliberately remove.
Here's where this gets personal.
Entry-level developer opportunities have plummeted ~67% since 2022. New graduates represent only 7% of Big Tech hires -- down from 32% in 2019. Over half of engineering leaders plan to hire fewer juniors because AI copilots let seniors handle more.
The logic is straightforward: why hire a junior for $90K when Copilot costs $10/month and a senior with AI tools can do the junior's work? The answer -- that juniors become seniors, and without juniors you have no pipeline -- is correct but doesn't show up in quarterly planning.
AI engineers earn 25% more than general tech roles. The average AI engineer salary jumped to $206,000 in 2025. AI-related job postings grew 74% year-over-year.
The market is bifurcating. On one side: developers who can architect systems, review AI output, and build production infrastructure. On the other: developers whose skills are indistinguishable from what AI can do.
Vibe coding puts you on the second side. Agentic engineering puts you on the first.
Only 29% of developers trust AI tools -- down 11 points from 2024. 84% are using them. That's a workforce that knows the tools are unreliable but uses them anyway because the productivity pressure is real.
The developers who'll succeed are the ones who channel that distrust into rigor. Not refusing to use AI. Not blindly accepting its output. Using it extensively while reviewing every line that matters.
One study cuts through the hype better than any other.
METR's randomized controlled trial tested 16 experienced open-source developers on 246 real issues, paying them $150/hour. The result: developers with AI tools took 19% longer than without them.
The kicker: those same developers predicted they'd be 24% faster. Even after the study, they believed they'd been 20% faster. The perception-reality gap is staggering. People feel faster with AI tools even when they're measurably slower.
Other studies show similar patterns:
| Study | Finding |
|---|---|
| METR (2025) | Experienced devs 19% slower with AI |
| METR update (2026) | Original devs still -18%; new recruits -4% |
| Faros.ai (10K devs) | 21% more tasks, but "no significant correlation with company-level improvements" |
| Google DORA (2024) | 7.2% decrease in delivery stability |
| Fastly (2025) | Experienced devs report 19% slower with AI editing overhead |
| Industry aggregate | Real productivity gains: 8-12%, not 10x |
The pattern: AI accelerates coding. But coding is a fraction of engineering time. Planning, understanding, debugging, reviewing, communicating -- AI doesn't accelerate those. It sometimes slows them down because you're now reviewing AI output on top of everything else.
Not all code deserves the same rigor. Here's my decision framework:
Vibe coding is the most dangerous idea in software engineering right now. Not because it doesn't work -- it works incredibly well for prototypes and demos. Because it teaches developers that understanding code is optional. And that lesson, internalized by a generation of new engineers, will cost the industry billions.
Here's what the data actually shows. AI-generated code has 2.74x more vulnerabilities. 45% of it contains OWASP Top 10 security flaws. It gets worse with iteration, not better. And the developers using it think they're faster when they're actually slower.
That combination -- invisible quality degradation plus false confidence -- is exactly how systemic failures happen. Not in one dramatic crash, but in thousands of small security holes, logic errors, and architectural decisions that compound over time.
The industry is already feeling it. 88% of developers report AI has negatively impacted technical debt. Analysts project $1.5 trillion in technical debt by 2027 from AI-generated code. CVE entries from AI code went from 6 in January to 35+ in March 2026. The curve is exponential.
Karpathy was right to evolve. His shift from "forget the code exists" to "there is an art and science and expertise to it" in exactly 12 months mirrors what every serious developer learns: AI coding tools are extraordinarily powerful if you understand what they're doing. The moment you stop understanding, they become liability generators.
The career implications are stark. Companies aren't hiring fewer developers because AI replaces them. They're hiring fewer junior developers (-67% since 2022) and paying more for senior developers who can review, architect, and correct AI output. The developers who embraced vibe coding and never learned the fundamentals are exactly the ones being squeezed.
I don't think vibe coding will disappear. It's too useful for prototyping. But I think the distinction between "person who vibe codes" and "engineer who uses AI agents" will become one of the most important career differentiators in tech. The first is a commodity. The second is increasingly rare, increasingly valuable, and increasingly well-compensated.
The question isn't whether you use AI to write code. Everyone does. The question is whether you understand what it wrote.