Vibe Coding

Confidence 0.95 · 22 sources · last confirmed 2026-06-20

The intuitive, exploratory mode of writing code with an AI agent — coined by Andrej Karpathy in 2024 and widely adopted as practitioner vocabulary through 2025–26. Karpathy’s working framing in the Sequoia AI Ascent interview (29 April 2026):

“Vibe coding is about raising the floor for everyone in terms of what they can do in software. The floor rises, everyone can vibe code anything, and that’s amazing — incredible.”

Vibe coding is the democratisation half of software-3.0: it makes software-creation accessible to people who don’t have classical software-engineering backgrounds. Its complement — agentic-engineering — is the professionals’ half, where the same substrate is wielded with quality discipline. The wiki holds both as two complementary modes of the same Software 3.0 substrate, not opposites.

Working definition

A vibe-coding session is one in which the human:

• Has a goal in mind (build a thing, try an idea, scratch an itch).
• Describes the goal to an agent in natural language, often without rigorous specification.
• Iterates by feel — looks at the output, asks for changes, accepts what works.
• Does not take responsibility for production-grade properties (security, scalability, maintainability, code aesthetics).

The vibe in vibe coding refers to the mood/feel-driven iteration loop, in contrast to the spec-and-test-driven iteration loop of traditional software engineering.

Key claims

The “December 2025 phase change” — vibe coding tipped from useful to trustworthy

Andrej Karpathy’s first-person account of the moment vibe coding stabilised:

“I’ve been using agentic tools like [Claude] code and adjacent things for a while, maybe over the last year as it came out, and it was very good at chunks of code and sometimes it would mess up and you have to edit them. December was this clear point where for me — I was on a break, so I had a bit more time, I think many other people were similar — and I just started to notice that with the latest models the chunks just came out fine. And then I kept asking for more and it just came out fine. And then I can’t remember the last time I corrected it. And I just trusted the system more and more. And then I was vibe coding.”

The structural claim: a single-month phase change in late 2025 when the correction overhead dropped to ~zero for a sufficiently broad class of tasks. This is not just “models got better” — it’s the threshold at which the iteration loop changes shape (from generate → review → edit → retry to generate → accept).

Two consequences Karpathy notes:

1. “My side projects folder is extremely full” — when the friction drops to zero, the volume of personal experimentation explodes.
2. The shift was not visible at the typical user’s distance — most people experienced AI in 2024 as “ChatGPT-adjacent thing.” Karpathy stresses on Twitter/X that you have to look again in late 2025 because the agentic-coherent-workflow has changed fundamentally.

Vibe coding raises the floor

The democratisation claim is the load-bearing operational outcome:

• People who could not previously write code can now build working software.
• Engineers who were intimidated by an unfamiliar stack (frontend if they’re backend; mobile if they’re web; etc.) can now ship across stacks.
• Side-project velocity for individual operators goes up multiple orders of magnitude.

The wiki’s enterprise-ai-adoption page tracks the organisational analogue of this; vibe coding is the individual analogue.

What vibe coding is not

Karpathy is explicit that vibe coding is not the same as professional software engineering at agent speed:

“Agentic engineering is about preserving the quality bar of what existed before in professional software. You’re not allowed to introduce vulnerabilities due to vibe coding. You’re still responsible for your software just as before.”

The danger in conflating them: vibe-coded artifacts shipped to production as if they were agentic-engineered carry risks (security, correctness, maintainability) that the vibe-coding loop didn’t validate. The wiki tracks this as an open hazard — see agentic-engineering for the quality-discipline complement.

When vibe coding is the right tool

• Side projects. Personal experiments, weekend builds, hobbyist apps. The Karpathy Menu Gen story is a vibe-coded artifact; he then reasoned about it through an agentic-engineering lens (and concluded most of it was spurious).
• Throwaway scripts. Glue code, one-off data transformations, ad-hoc analysis.
• Proof-of-concept demos. Where the goal is to learn whether something is possible, not yet to ship it reliably.
• Education and exploration. Learning a new framework or library by having an agent walk you through a working example.
• Stakeholder-influence prototypes for non-engineers. Per Nika 2025, the canonical PM use of v0 / Lovable / Google AI Studio is to turn a PRD into a clickable prototype before asking for engineering allocation, then take the prototype into a product review where stakeholders can touch / feel / move-around the vision. The prototype is a communication and influence-engineering tool, not a design artifact. This is vibe coding’s first-class non-engineer professional use — the floor-raising claim made operational for the PM audience specifically.

When vibe coding is the wrong tool

• Production systems handling third-party data, money, or identity. Karpathy’s email-matching-Stripe-and-Google-accounts bug is a vibe-coding-shipped-to-production failure mode.
• Code that other engineers will maintain. Vibe-coded code is often bloaty, copy-paste-heavy, awkward in its abstractions, brittle (Karpathy’s own description) — fine for a one-off, expensive to inherit.
• Anything that requires correctness verification. Vibe coding’s iteration loop trusts the model’s output; for verifiable domains, the more rigorous agentic-engineering loop (with formal contracts and adversarial testing) is the right pattern.
• Anything safety-critical. Same reason as above — the vibe loop doesn’t validate against an explicit quality bar.

Convergence with other wiki claims

Source	What it adds to vibe coding
Karpathy 2026	Original coinage + December 2025 phase change + the floor-vs-ceiling contrast with agentic-engineering
Dell’Acqua et al. 2026	The empirical underbelly: vibe-coded output inside the jagged-frontier is genuinely good; outside it can degrade performance even relative to the un-AI-assisted baseline
AEI 5	Capacity-ceiling caveat: vibe coding works at single-task scale; once tasks compose (multi-step, long-horizon) the ceiling becomes load-bearing, and you need agentic-engineering
Nika 2025 (How I AI)	The PM-side worked example + the earliest contemporary primary source for the December 2025 phase change. A Google AI Product Lead, on camera in late November / early December 2025, doing the workflow live (Perplexity → custom GPT → v0 → Veo/Sora → NotebookLM) and treating the tool stack as already-trustworthy. Substantiates floor-raising with a non-engineer professional (PM) audience: vibe coding’s clickable-prototype pattern as stakeholder influence engineering, not just hobbyist iteration. “PMs that use AI are the ones that are going to take over the role of people that don’t use AI.”
Böckeler 2026 (Thoughtworks at QCon London)	Conference-circuit timeline anchor. Böckeler’s year-on-year QCon retrospective dates “the term vibe coding was about 2 months old” at QCon London 2025, anchoring the term’s coinage to early 2025 from an independent observer (consultancy practitioner). Frames vibe coding as one end of an autonomy spectrum the broader agent-harness discipline now attempts to widen safely — directly compatible with Karpathy’s floor-vs-ceiling split
The Daily)	75-developer field-report anchor. The wiki’s first cross-industry survey of working software developers — Thompson reports “a majority outsourcing day-to-day programming to AI”, with small startups at ~100% AI-written code (full-day features now ~30 min, ~20x faster) and Google at ~40-50% AI-written code with ~10% overall speedup. Dates the developer-side phase change to “last 6 months, accelerating in the last 3” — i.e. November 2025 → February 2026, ~4 months behind Nika’s PM-side December-2025 phase change. Vibe coding observed in production at the small-startup floor; the consultant-observer + journalist-observer + vendor-engineer + practitioner vantages now all align on the same transition
Jassy 2025 (Amazon CEO on HBR IdeaCast)	Pre-coinage CEO anchor for the floor-raising thesis — the wiki’s earliest vendor-CEO statement of the natural-language-coding democratisation prediction. Jassy in May 2025: “the number of people who are going to be able to be software developers is going to go up exponentially. Because you’re going to have these coding apps that allow you to use natural language to describe what you want to go build. And it’s going to be very empowering.” Predates Andrej Karpathy’s vibe-coding coinage by ~12 months and Nika’s December-2025 PM-side worked example by ~7 months. The qualifying “but” clause (US-education-quality concern: “30 out of 35 developed countries in efficacy of education”) anticipates the wiki’s ai-deskilling / durable-skills concern that floor-raising democratisation depends on a baseline of human capability that may itself be eroding
Startup School)	Company-scale extension of the floor-raising thesis. Hu: “Everyone comes to meetings with working prototypes not pitch decks. Everyone builds and ops support sales.” Vibe-coding as an organisational norm rather than just an individual skill — extends Nika’s PM-side worked example into a cross-function expectation. The startup-edge claim — “design your systems, workflows and culture around AI from the start and operate thousand times faster than the incumbents” — applies the floor-raising thesis at the firm-formation level. Adds the YC-partner observer vantage to the wiki’s growing cluster on the prediction-and-its-materialisation arc (Jassy May 2025 prediction / Nika December 2025 worked-example / Karpathy April 2026 coinage / Thompson April 2026 75-developer survey / Hu April 2026 company-scale prescription)
YC Founder Firesides)	Vendor-CEO platform-launch anchor + audience refinement. Masad: “September 2024, Replit became the first like what’s called vibe-coding product where we abstracted away code entirely.” Replit’s commercial vibe-coding launch predates Karpathy’s December-2025 phase-change framing by ~3 months on the platform side — the floor-raising transition required infrastructure to materialise, and Replit shipped that infrastructure before the broader phase change. Two further refinements: (a) Traditional developers don’t get the most value — “developers like the pain… It’s sort of like a craftsperson liking to build their own tools. People that are getting the most value tend to be the more tech-adjacent ones — PMs, designers, entrepreneurs”; the audience-pipeline framing now generalises beyond Nika’s PM-side anchor to all tech-adjacent non-developers + domain experts (physical therapists, pool-business operators, sports-club ops, parents tracking rare conditions). (b) Post-prompting product direction — “I actually think we’re headed to a post-prompting world… at some point Agent 5 or maybe sooner, you should be able to tell Replit, ‘every day build me a SaaS company and try to market it and see what works.‘” Marks the wiki’s first identification of a post-prompting trajectory inside vibe-coding-as-platform. Also names the vibe-coding-resident team — a generalist team with deliberately vague mission “go around the company and make it better” deputising agents — as an organisational pattern materialised inside Replit itself. Pair with Hu’s “everyone builds and ops support sales” (the day before): same idea, founder-CEO worked example
GStack)	Live demo of vibe-coding-with-discipline at YC-president scale. Tan walks through idea → adversarial review → design mockups → code → QA in a single Claude Code session — the entire ~10-min demo is a worked example of the floor-raising thesis materialised. Closing peroration: “the barrier to building just collapsed. The only question left is what are you going to build? It’s time to let it rip. Go make something people want.” GStack as the packaged toolkit that turns the floor-raising-into-a-discipline transition into a one-CLI-install workflow. With Hu the day before, the YC-batch-context anchor pair (Tan 23 April + Hu 24 April) frames vibe-coding-as-organisational-baseline from the platform that funds the floor-raised cohort
Anthropic Applied AI interview)	Strongest 2026 named-company anchor on the floor-raising thesis materialised at scale. 7M users in 190 countries; 70–80% have never written a line of code; $100M ARR in 8 months. The audience refinement Masad named at Replit (tech-adjacent non-developers + domain experts) is operationalised here as small businesses and domain experts as the primary user class — the wiki’s clearest 2026 vignette is Christy, a clinical psychologist and equestrian coach in Alaska, who had paid $50K to a Nova Scotia dev shop for an app that didn’t work, switched to Emergent and shipped “Equ” on the App Store with hundreds of users. Jha’s framing of zero marginal cost serving million niches of SMB / domain-expert needs (“70% of employment, 50% of GDP”) is the wiki’s compactest statement of vibe-coding-as-economic-democratisation. With Replit, the wiki now holds two named-CEO 2026 anchors for the natural-language-to-production-app platform category — distinct technical architectures (Replit on own Agent N stack; Emergent on Claude + own multi-agent harness on Kubernetes), shared user-class thesis
Figma Make customer story 2026	Incumbent-design-tool-vendor expression of the floor-raising thesis. “I don’t code. One of the best innovations of the past couple of years is that I really don’t have to in order to create the things I want to do.” Same first-person “I don’t have to code” claim as Karpathy / Nika / Replit-users but from inside a design-tool incumbent, with the discriminator “every single person who has taste can just enact it all that easier” — taste as the load-bearing human contribution, Claude as the enactment layer. Adds the design-tool-vendor vantage to the wiki’s vibe-coding anchors alongside Replit / Emergent (general-purpose vibe-coding platforms) and the YC-batch context (Tan / Hu)
Microsoft Dec 2025	The “can you vibe debug it?” security caveat, with a worked failure. Randell: “you can vibe code, but can you vibe debug it?” — then narrates vibe-coding a Chrome bookmark-manager extension with Claude, having Codex review Claude’s code (“looked good”), pushing to GitHub — and GitHub Advanced Security still catching a URL-injection vulnerability multiple times. “Two smart AI agents still screwed up and GitHub had my back.” The wiki’s cleanest vendor-keynote articulation of the vibe-coding-needs-a-safety-net corollary: floor-raising is real, but shipping vibe-coded output to production without scanning/review is the failure mode — directly compatible with the floor-vs-ceiling split and the responsible-ai human-in-the-loop default
Google DeepMind)	Platform-scale floor-raising datum + the personal-software thesis. AI Studio saw 350,000 Android apps built in one week — “apps that probably no one was going to build before,” many personal/single-user. Kilpatrick: “building software to solve your personal problem is very real right now.” Most-built AI Studio categories: ~20% finance/crypto, games (was ~20%), personal productivity, gen media. Extends the Emergent/Replit domain-expert audience claim onto the Android-app-on-AI-Studio surface, with Android framed as “the platform for builders.” The same interview’s model-eats-the-harness thesis also reinforces the agent-harness note below: vibe coding runs on a thin/absent harness because the model absorbs the scaffolding
YC Lightcone 2026	The “anti-slop” opinionated-vibe-coding-platform anchor + domain-expert audience. Bryant Chou (Webflow → Ploy) builds vibe-coded award-winning sites but explicitly frames Ploy as the anti-slop counter to generic vibe-coding output: a deterministic “design slurper” that extracts a consistent design system from an existing site (no 10 header-font variants) plus a curated lookbook of ~3,500 web-design prompts that gives the model inspiration, not copies — “if it was AI slop, that wouldn’t have worked.” The Andy-Warhol-factory framing (models are “the factories for human creativity”) and the “you need an opinion about how websites should be made” line make the floor-vs-ceiling split concrete inside a vibe-coding product: the platform supplies opinionated discipline so non-experts clear the quality bar. Extends the Replit/Emergent domain-expert audience claim and pairs the vibe-coding floor-raising thesis with the expertise differentiator
O 2026	The corporate-research ceiling-side negative claim. “Our top engineers are not vibe coding. They are not prompting for code. What they’re doing is designing environments, setting the guardrails, creating the systems so that agents and humans can work together toward a goal and a broader purpose.” The cleanest Google-Research articulation of the floor-vs-ceiling split — vibe coding raises the floor (acknowledged elsewhere in the talk via the Search-org PM story); but at the ceiling, top engineers explicitly do not vibe code, they practise agentic-engineering. With DORA’s amplifier-and-mirror framing: “if you lack deep engineering understanding… AI can run the risk of generating technical and cognitive debt at an incredibly fast rate” — the published-by-Google operationalisation of the “why vibe-coding-shipped-to-production is the wrong tool” claim

| Snowflake (McKinsey Podcast) 2026 | The floor-raising thesis inside an enterprise, from the CEO chair. Ramaswamy’s sales reps “don’t think twice about shipping a new visualization dashboard or honestly a new application … all they have to do is describe what they want done in English.” He frames the conversational interface as the new natural interface for software — replacing “bespoke web interfaces … stylized to serve the masses” — and the learning loop as ask → explain → ask-differently → debate. Pairs the non-engineer-shipping-software claim (Replit/Emergent/Figma vantages) with a public-data-company-CEO vantage; the floor-raising and the ceiling (“uber programmers 50–100×”) are named in the same breath |

| AI Native DevCon 2026 | Vibe coding as an enabler of the group-tech-lead operating mode (from the coiner of harness engineering). “I think vibe coding is a big part of what it takes to be successful here, because there’s a ton of guardrails that only affect my local development process. This code can be gross, but it brings into possibility this idea that I don’t need to care about some parts of the software production function.” The payoff: operating “like a group tech lead or an org lead where I don’t have visibility into every single engineer’s activity on the keyboard” — caring about invariants/interfaces, not keystrokes. He also vibes guardrails into place (100% coverage + table-driven tests, migrate the codebase in one go) and vibes up an Xvfb+ffmpeg rig to record reproduction videos. The clearest statement that the floor-raising mode is a load-bearing tool inside the ceiling discipline, not opposed to it |

Debates and supersession

• Is “vibe coding” a derogatory term or a legitimate mode? When the term first spread on Twitter/X in 2024–25, some practitioners used it dismissively (“that’s just vibe coding, not real engineering”). Karpathy’s Sequoia framing resolves this: vibe coding is a legitimate mode (floor-raising democratisation) — the dismissive framing is misplaced because it conflates vibe coding with agentic engineering. The wiki adopts Karpathy’s resolution.
• Does the December 2025 phase change generalise beyond Karpathy’s experience? No second source has yet ingested a parallel first-person account. Partially answered by Nika 2025 — a Google AI Product Lead doing the workflow live on camera in late November / December 2025, contemporary with Karpathy’s named pivot. This isn’t a survey or controlled study, but it is a second independent practitioner whose tool-stack is already trustworthy by the date Karpathy retroactively names. The remaining open question is whether a population-level signal (survey / empirical study) confirms it across professional audiences; the wiki now holds two practitioner anchors at the engineer (Karpathy) and PM (Nika) ends.
• Will vibe coding always need an agentic-engineering complement, or will model improvements close the gap? Karpathy’s view: “I’m hoping that improves… there’s nothing fundamental preventing it. It’s just the labs haven’t done it yet.” If labs add aesthetics-as-RL-reward and agent code stops being bloaty / awkward / brittle, the gap between vibe coding and agentic engineering narrows substantially. Open question.

Related concepts

• software-3.0 — the substrate; vibe coding is the floor-raising mode of Software 3.0.
• agentic-engineering — the discipline complement; same substrate, professional quality bar.
• ai-agents — the intern entities you vibe-code with.
• agent-harness — vibe coding can happen with a thin or absent harness; agentic-engineering requires a substantial one.
• jagged-frontier — vibe coding inside the frontier is genuinely good; outside it, the iteration loop trusts model output that may be wrong.

Open questions

• Empirical measurement of the December 2025 phase change. Karpathy’s account is first-person; a population-level signal would lift this concept’s confidence.
• The vibe-coding-to-agentic-engineering boundary in practice. When does a vibe-coded side project earn promotion to agentic engineering? Karpathy’s Menu Gen example was vibe-coded then replaced entirely by a Software 3.0 prompt — a rarer pattern than gradual hardening. Worth tracking which trajectory dominates.
• Will vibe coding remain the individual mode or become the team mode too? Currently it’s framed for solo operators. If teams adopt vibe-coding loops collectively (with each member trusting the agent’s output without review), the safety implications change substantially. Open.
• Tooling specialisation. Will vibe-coding-specific IDEs/tools emerge that are explicitly not agentic-engineering-grade (lower friction, weaker guardrails)? The pet-vibe-coding tools that exist today (Replit’s Agent, Cursor’s vibe modes) hint at this; second sources documenting the trajectory would be useful.