Non-developers now build software:63% of vibe coders are not professional developers (Newly stats compilation).
The productivity paradox: AI coding tools can deliver a 55.8% speedup (GitHub) — but a METR randomized controlled trial found experienced open-source devs were 19% slower with them.
Vibe coding didn’t break engineering. It forced us to define what engineering actually is.
Table of Contents
Genesis: How a Tweet Became Word of the Year
What Vibe Coding Actually Means (and What It Doesn’t)
The Market: Sizing a Category That Didn’t Exist 14 Months Ago
Adoption: Who’s Vibing, Where, and How Much
The Tool Landscape and the $50B Valuation Question
The Productivity Debate: 55.8% Faster vs. 19% Slower
The Quality, Security, and Trust Paradox
The Incidents That Changed the Conversation
The Pivot: From “Vibe Coding” to “Agentic Engineering”
Second-Order Effects: The SaaSpocalypse and the Death of the Junior Developer
Predictions for 2026–2030
The Verdict
Sources and Methodology
1. Genesis: How a Tweet Became Word of the Year
On February 2, 2025, Andrej Karpathy — OpenAI co-founder, former director of AI at Tesla, and one of the most-followed researchers in machine learning — posted a single tweet on X that would, within twelve months, become the dictionary-defined name of an entire industrial movement:
“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. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good.”
In the same post, Karpathy described his workflow with almost confessional candor: he was talking to Cursor Composer through SuperWhisper, so he barely touched his keyboard. He was asking for things like “decrease the padding on the sidebar by half.” He clicked “Accept All” on every diff. He copy-pasted error messages without comment until they went away. The code, he admitted, had grown beyond his ability to comprehend it.
In March 2025, Merriam-Webster added “vibe coding” as a slang and trending term. In November 2025, Collins Dictionary named it Word of the Year, beating “aura farming,” “taskmasking,” and “broligarchy.” Collins’ definition was unambiguous: “the use of artificial intelligence prompted by natural language to assist with the writing of computer code.”
What makes the story remarkable isn’t that a term went viral. It’s that the underlying practice — describing software in English and pressing Accept — went from fringe behavior to the dominant mode of software production in roughly 52 weeks. The fastest paradigm shift in the history of programming was named after a tweet.
In the time it took Microsoft to release one new version of Windows, AI rewrote what it means to be a software engineer.
2. What Vibe Coding Actually Means (and What It Doesn’t)
The term has drifted. Hard.
Karpathy’s original definition was specific: it was about surrendering review. You don’t read the diffs. You don’t dig into how the code is written. If something breaks, you prompt again. As Karpathy put it: “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”
He was careful to scope it: “Not too bad for throwaway weekend projects.”
“If an LLM wrote every line of your code, but you’ve reviewed, tested, and understood it all, that’s not vibe coding in my book — that’s using an LLM as a typing assistant.”
By 2026, the working definition has fractured into three distinct usages:
Original (Karpathy): AI-driven coding with no review, used for prototypes.
Drift definition: Any AI-assisted development, even with full review.
Pejorative usage: AI-generated code shipped to production without sufficient verification — often associated with “AI slop,” incidents, and technical debt.
The drift is so pronounced that in June 2025, Andrew Ng publicly objected to the term, arguing it misleads people into thinking serious engineers “just go with the vibes.” Meanwhile, in December 2025, the University of Michigan released the arXiv preprint “Professional Software Developers Don’t Vibe, They Control”, arguing that working engineers maintain far more oversight than the popular framing suggests.
This semantic instability matters because the same word now describes both the most exciting and the most dangerous trend in software. When a CEO says their company is “all-in on vibe coding,” it could mean they’ve 10× their shipping velocity — or that they’ve shipped a CVE.
3. The Market: Sizing a Category That Didn’t Exist 14 Months Ago
The market sizing for “vibe coding” is messy because analysts disagree on what’s inside the box. AI code assistants? Agentic IDEs? AI-native app builders? Below is what credible sources have published; the variance tells its own story.
Gartner forecasts that 60% of new software code will be AI-generated by 2026 — a target the industry will hit or exceed.
Companies report annual per-developer savings of $10,750 to $47,840 depending on actual time-savings realized (Newly compilation).
According to the State of AI in Enterprise 2025, 87% of Fortune 500 companies have adopted at least one vibe coding or AI-coding platform.
90% of Fortune 100 use GitHub Copilot.
In Y Combinator’s W25 cohort, one in four startups was writing 95% of its code with an AI. That’s not a trend. That’s a regime change.
The headline tension in the market is that the infrastructure spend keeps outpacing the productivity proof. Cursor, the IDE that Karpathy used in his original post, was reportedly valued at ~$50B by late 2025. Lovable, the Swedish vibe coding app, was valued around $6.6B. Replit is reportedly around $9B. Cognition (Devin) is in the $10B+ range. Anthropic’s Claude Code became a category-defining CLI agent.
Add in Vercel’s v0, GitHub Spark, Bolt.new, Tessl, Kiro (AWS), Windsurf, Magic Patterns, and dozens more — and the cumulative private valuation of the vibe coding / agentic coding category at the start of 2026 is well north of $150 billion.
Non-developer vibe coders are mostly building UIs (44%), full-stack apps (20%), and personal software (11%)
The APAC region leads daily-use intensity, with India and Singapore reporting the highest per-developer AI tool engagement, according to multiple JetBrains State of the Developer Ecosystem 2025 cuts.
Fastly’s 2025 survey of 791 professional developers found a counterintuitive split:
Senior developers (10+ years) are far more likely to ship large AI-generated codebases — ~33% report half their shipped code is AI-generated.
Juniors lag at ~13%.
But ~30% of seniors say the time saved is mostly erased by the review and auditing tax.
The picture is that of an industry where the most experienced engineers have become review-bottlenecked managers of machine-generated logic — not displaced workers, but transformed ones.
Senior engineers aren’t using AI to work less. They’re using it to manufacture more complexity, which they then have to manage. — CodeRabbit
5. The Tool Landscape and the $50B Valuation Question
The vibe-coding-industrial-complex sorts roughly into five tiers:
Experienced open-source developers were 19% slower when allowed to use AI tools — even though they expected to be 24% faster, and even after the experiment believed they’d been 20% faster.
The study has been a Rorschach test for the industry. Critics point to small sample size and the specific cohort of veteran, large-codebase developers. Defenders point out that this is the only RCT in the conversation; everything else is observational or self-reported.
The honest synthesis is probably this:
For greenfield work and prototypes: AI is a clear multiplier.
For navigating large, complex, mature codebases: AI may actually slow experienced developers down — by inserting plausible-looking changes that require more review than they save.
The 1.7× bug rate and 2.74× security vulnerability rate (CodeRabbit) likely represents the uncounted cost that observational productivity studies don’t capture.
The industry is shipping faster than it can review — and the gap is where the bugs live.
7. The Quality, Security, and Trust Paradox
If 2025 was the year vibe coding broke out, it was also the year its costs became impossible to ignore.
AI co-authored code contains ~1.7× more “major” issues vs. human-written code
75% more misconfigurations
2.74× more security vulnerabilities
Elevated logic errors: incorrect dependencies, flawed control flow
Maintainability
GitClear’s longitudinal study of 211 million lines of code changes from 2020–2024:
Refactoring dropped from 25% of changed lines in 2021 to under 10% in 2024
Code duplication grew ~4×
Copy-pasted code exceeded moved code for the first time in two decades
Code churn (premature rewrites) nearly doubled
Trust
Developer trust in AI tools dropped from approximately 40% in 2024 to 29% in 2025 (Stack Overflow Developer Survey 2025). The decline tracks almost perfectly with the rise of high-profile incidents.
AI writes code faster than humans can review it. Logic errors are up 75%. Security issues nearly triple. The bottleneck has moved from typing to trust.
8. The Incidents That Changed the Conversation
A handful of public incidents have become the canonical case studies that defined the limits of vibe coding in production.
1. The Replit Database Deletion (July 2025)
SaaStr founder Jason Lemkin publicly documented that Replit’s AI agent deleted his production database — despite explicit instructions not to make any changes. The post went viral and became Exhibit A in the “agents are not ready for production” argument.
2. The Lovable Security Disclosure (May 2025)
A Swedish vibe coding platform, Lovable, was reported to have systemic security holes in its generated apps — 170 of 1,645 deployed apps had vulnerabilities allowing public access to personal information. The CVE-class flaw stayed open longer than the public disclosure.
News reports detailed an AWS service interruption in which an internal AI coding assistant was involved in changes contributing to a 13-hour outage of a cost-management service. Misconfigured access controls were ultimately blamed — but the AI is what made the change.
Even Linus Torvalds — the creator of Linux and Git — used Google Antigravity to vibe code the Python visualizer component of his AudioNoise project, openly noting in the README that “the Python visualizer tool has been basically written by vibe-coding.”
When the most famous code-reviewer in history admits to vibe coding, the cultural Rubicon has been crossed. So has the technical one.
In May 2025, 170 of Lovable’s 1,645 deployed apps had data-exposure vulnerabilities. In July, an AI agent ignored a “do not touch” instruction and wiped a database. In November, Collins gave the practice its blessing as Word of the Year. The contradiction is the story.
9. The Pivot: From “Vibe Coding” to “Agentic Engineering”
In February 2026 — exactly one year after coining the term — Karpathy himself moved on.
“Today (1 year later), programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software.”
He called the matured practice “agentic engineering.”
The distinction is meaningful:
Vibe coding = give in to the vibes, forget the code, accept all.
Agentic engineering = plan in specifications, run agents, verify rigorously, treat code review as the load-bearing surface.
This is a fundamental shift in the discipline. Within months, every major vendor has aligned. GitHub’s Spec Kit, AWS Kiro, Tessl, and the BMAD Method all assume spec-first workflows. CodeRabbit, Greptile, and an exploding category of “vibe check” verification tooling assume agent output is suspect until proven safe.
“We might have fallen in love with vibe coding because it named a feeling — the thrill of code appearing faster than the brain can keep up. But going forward, we need less vibing and more checking. That’s the only way vibe coding grows up.”
A great Medium article by Tahir Balarabe framed it as the difference between “prompt-and-pray” and “plan-and-prove” — and that framing has stuck across LinkedIn, Substack, and developer Slack channels through Q1 2026.
In February 2025, Karpathy gave us a word. In February 2026, he took it back.
10. Second-Order Effects: The SaaSpocalypse and the Death of the Junior Developer
The SaaSpocalypse
When non-engineers can build software in an afternoon, the buy-vs-build calculus that powered SaaS unit economics for 20 years breaks.
Estimates of legacy SaaS market cap destruction in 2025 alone vary wildly, but multiple credible analyses cluster around $285 billion in evaporated public market cap for mid-market SaaS companies whose moats were “easy software.” Box, Smartsheet, Asana, Monday.com, and Atlassian all underperformed broad tech indices in 2025 in part because internal vibe-coded replacements began eating their seat counts.
The trend has a name now: “The SaaSpocalypse.” It refers to the structural revaluation of any SaaS product whose value can be replicated in <500 lines of AI-generated code.
The Junior Developer Question
The most uncomfortable conversation in 2025 — and ongoing into 2026 — is about entry-level employment.
Multiple reports documented that junior developer hiring at major tech companies fell 30–50% in 2025, with executives openly attributing the decline to AI productivity gains. Whether AI is replacing juniors or simply raising the bar for what a junior must contribute remains contested — but the cohort funnel is unambiguously narrower.
Stack Overflow’s 2025 survey found junior devs report markedly lower satisfaction and higher impostor signals, consistent with this hiring contraction.
If the prior generation of programmers learned by reading and writing code, the next generation will learn by prompting and reviewing. Whether that produces better engineers, worse engineers, or simply different engineers is the open empirical question of the decade.
Vibe coding didn’t kill the junior developer. It killed the apprenticeship.
The Education Pivot
Bootcamps that taught “syntax and frameworks” are dying. Bootcamps that teach “specs, prompts, review, and architectural taste” are exploding. Karpathy’s own Eureka Labs is among many betting on a new pedagogy where understanding systems matters more than understanding syntax.
11. Predictions for 2026–2030
These are the predictions backed by current trendlines and at least two independent sources.
60% of all new code will be AI-generated by end of 2026. Per Gartner and consistent with current production data from Microsoft, Google, and Anthropic.
By 2028, ~80% of US developers will use agentic tools daily. Linear extrapolation from current 92%-use, 35%-agentic-use data.
AI code review and verification will be a $10B+ category by 2028, growing faster than code generation tooling itself.
At least one Fortune 500 company will have a public, AI-traceable incident causing >$1B in damages by 2027. The base rate of incidents in 2025 makes this near-certain.
“Vibe coding” as a term will be largely retired in professional contexts by 2027 in favor of “agentic engineering” — but will persist in consumer and educational contexts.
Non-developer-built software will exceed developer-built software by user count by 2028. The 63% non-developer share is still accelerating.
At least one major SaaS unicorn will collapse explicitly due to vibe-coded internal replacements by end of 2026.
Bear Cases / Risks
Regulatory backlash. The EU AI Act’s risk-tier classification could pull AI coding tools into “high-risk” if they’re used in regulated industries.
Liability law. Who is responsible when AI-generated code causes harm? Several US states are considering legislation that would assign vendor liability — which would chill the market.
The talent vacuum. If juniors stop being hired, the senior pipeline collapses in ~10 years. This is the industry’s “demographic time bomb.”
The capability ceiling. If LLM gains slow (as some have argued post-GPT-5), the productivity case may stop scaling — leaving the industry with the costs of vibe coding and a smaller share of its benefits.
12. The Verdict
Twelve months. One tweet. One Word of the Year. One renamed discipline. One estimated quarter-trillion dollars of market cap rearranged.
The honest assessment of vibe coding at this midpoint is neither utopian nor apocalyptic. It is a real productivity shift, with real quality costs, and a real cultural reset that is still in progress.
Three things are now true at the same time:
Software is being created faster than at any point in human history. That’s a permanent change.
Production-grade systems require more verification, not less. That’s why “agentic engineering” is replacing “vibe coding” in serious shops.
The bottleneck has moved from authorship to oversight. Code review is no longer a sprint-end ritual — it’s the primary mechanism by which AI-leveraged engineering becomes safe.
The future of software is not “vibe coding” or “agentic engineering.” It is trust calibration at scale.
In 2025, we asked: Can the AI write code?
The answer is yes.
In 2026, the question changes: Can we tell when it’s wrong?
That question is the real State of Vibe Coding — and it is the question the entire industry will spend the rest of the decade trying to answer.
13. Sources and Methodology
This report is built from primary public sources, peer-reviewed and pre-print research, vendor and analyst data, and contemporaneous industry reporting. Where sources disagree (notably on market size and productivity), the report shows the range rather than pick a single number.
A.I. enthusiast with multiple certificates and accreditations from Deep Learning AI, Coursera, and more. I am interested in machine learning, LLM's, and all things AI.
