From vibe coding to agentic engineering
"Vibe coding" had a remarkable eighteen months: from a tweet to a dictionary entry to a way of working that millions tried. Then the limits showed up, on schedule. The interesting story in 2026 isn't that vibe coding failed — it's where the durable value moved once the hype settled. Productivity claims in this area are genuinely mixed, so we've flagged the contested ones rather than smoothing them over.
What vibe coding got right
Andrej Karpathy coined "vibe coding" in February 2025 — leaning fully into prompting, barely reading the code, accepting suggestions and iterating by feel. It went properly mainstream: Merriam-Webster added the term, and Collins named it word of the year for 2025.
What it got right is real and shouldn't be dismissed: the cost of a first draft collapsed. Prototyping something to see if it's worth building is now nearly free in time and effort.
The app-builder market is the proof. Lovable, Bolt, v0, and Replit Agent all scaled fast. Lovable reportedly reached a ~$400M annualized run-rate and 100k+ projects per day by early 2026 (company-stated and third-party estimates). Cursor reportedly went from ~$100M to ~$2B ARR (estimates). Read those numbers as directional — they're company-stated or estimated, not audited — but the direction is unambiguous: a lot of people found real value in fast, AI-driven first drafts.
The three-month tech-debt wall
The catch is now well-documented. Vibe-coded projects accumulate tech debt fast — some sources cite roughly 3× faster than conventional code — and many teams report a "three-month wall" where the quick prototype becomes unmaintainable and stalls.
The quality data is genuinely mixed, and worth stating honestly:
- Veracode-style security audits report that a large share of AI-generated code carries vulnerabilities, with figures ranging roughly 40–62% depending on the study and methodology. That's a wide band — which is itself the point. Don't treat any single figure as settled.
- A METR randomized controlled trial (2025) found that experienced open-source developers were about 19% slower with AI on their own repositories — while feeling about 20% faster. The perception of speed and the measured reality diverged.
None of this means AI coding doesn't help. It means productivity claims deserve caution, and "it feels faster" is not evidence that it is. For unfamiliar greenfield work the gains look real; on mature, well-understood codebases the picture is much murkier.
Spec-driven + agentic engineering
The field's response was to put structure back in. The move is from pure "vibes" to spec-driven development: write the requirements and design first, then generate under supervision against that spec. You're still using AI heavily — you're just not flying blind.
Karpathy himself moved past pure "vibes": around the one-year mark he proposed the term agentic engineering — you're "orchestrating agents... and acting as oversight" rather than writing the code directly. The framing matters: it's not a retreat from AI, it's a maturation of how it's used. In parallel, enterprises started adding governance around AI-generated code — staging environments, mandatory review, audit trails — the same controls any other production input gets.
What this means for an agency
If the first draft is cheap and commoditized, then the value migrates to everything the first draft doesn't give you. The developer's job shifts from writing code to orchestrating agents and owning judgment: system design, testing, accessibility, security, and design-system governance.
So the durable agency offer in 2026 isn't "we'll vibe-code your app." It's production-hardening, governance, and long-term support of AI-generated codebases — taking the prototype across the three-month wall and keeping it maintainable.
Concretely, the way we work it: use AI builders for prototypes and pitches, where speed is the whole value — then ship the production build by hand, with review and tests. The prototype proves the idea cheaply. The hardened build is what you actually charge for, and what survives.
The mental model that's held up for us: AI gets you to a working draft fast, but a draft is not a product. The distance between the two — tests, accessibility, security review, a design system that doesn't drift, code someone can still change in six months — is exactly the distance that didn't get cheaper. That gap is the work, and it's where a studio earns its keep.
Sources: Karpathy's original "vibe coding" post (Feb 2025) and his later "agentic engineering" framing (Feb 2026); Merriam-Webster (added the term, Mar 2025) and Collins (Word of the Year 2025); company-stated and third-party figures for Lovable and Cursor; Veracode (2025) and CSA (2025) AI-code security audits; the METR 2025 RCT. Figures are as of mid-2026 — not audited. The ~3× tech-debt multiplier is the softest (directional, conditional on QA practices); the METR ~19% slowdown / ~20% perceived speedup and the 40–62% vulnerability range come from the named studies.