Agentic Engineering: Karpathy's New Term and What It Means for Non-Developers

How emerging AI infrastructure creates new opportunities for service businesses.

Karpathy coined vibe coding in 2025. In February 2026 he replaced it with agentic engineering. Here is what changed, why it matters, and what it means for your business.

In February 2025, Andrej Karpathy — co-founder of OpenAI and former head of AI at Tesla — coined "vibe coding" to describe the practice of using AI to generate code from natural language prompts without closely reviewing the output. It became the Collins Dictionary Word of the Year.

Exactly one year later, in February 2026, Karpathy declared vibe coding passé. His replacement term: "agentic engineering."

The shift isn't semantic. It reflects a fundamental change in how AI-assisted development works — and it matters for anyone building software products, whether you're technical or not.

What changed

When Karpathy coined vibe coding, AI coding tools were limited. They could generate snippets, complete functions, and help with simple tasks. The "vibe" approach — prompt, accept, run it, see if it works — was appropriate for that capability level.

A year later, AI tools are dramatically more capable. They can plan multi-step implementations, execute across files, run tests, debug issues, and iterate autonomously. The human role shifted from "writing code with AI assistance" to "directing AI agents that write the code."

As Karpathy wrote: "Today, programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny."

Agentic engineering vs vibe coding

The distinction is clear when you look at the workflow:

Vibe coding: You prompt. The AI generates code. You accept it without deep review. If it works, you ship it. If it doesn't, you paste the error back and try again.

Agentic engineering: You start with a plan — architecture decisions, data model, security approach. You break work into well-defined tasks. AI agents execute each task. You review every output with the same rigor you'd apply to a human developer's work. Testing is extensive.

As Addy Osmani put it: vibe coding is YOLO. Agentic engineering is AI doing the implementation while the human owns architecture, quality, and correctness.

The data proves why this distinction matters

A randomised controlled trial by METR — the most rigorous study of AI coding productivity to date — found that experienced developers using AI tools took 19% longer to complete tasks than those working without AI. Before starting, developers predicted AI would make them 24% faster. After finishing (and being measurably slower), they still believed AI had sped them up by 20%.

This is what vibe coding produces: a perception of speed without the reality of it. Developers spent time reviewing, testing, and rejecting AI-generated code — less than 44% of suggestions were accepted — while context-switching between "coding mode" and "prompting mode" disrupted the deep focus that makes experienced developers fast.

Agentic engineering addresses this directly. Instead of accepting every suggestion and hoping it works, you plan the work, give AI agents well-scoped tasks, and review output with the same rigour you'd apply to a junior developer's pull request. The structure eliminates the perception gap.

Spotify: agentic engineering at scale

The clearest example of agentic engineering in production emerged in February 2026. Spotify co-CEO Gustav Söderström revealed that the company's best developers "have not written a single line of code since December." They shipped 50+ features in 2025, hit 751 million users, and posted record operating income of €701 million.

Spotify built an internal system called Honk — layered on top of Anthropic's Claude Code — fine-tuned to their specific codebase, architectural patterns, and development practices. Engineers now instruct AI agents, review output, approve changes, and merge to production. Some deploy features from Slack on their morning commute.

This is agentic engineering, not vibe coding. The engineers still need to understand what "correct" looks like. They need architectural knowledge, quality standards, and product judgment. They shifted from typing code to directing agents — and the key detail everyone misses is that the shift only works because they have years of deep codebase knowledge. Strip away that experience and you've got someone approving code they can't evaluate.

The Axios CTO told a similar story: an engineer used AI agent teams to complete in 37 minutes what previously took three weeks. But what compressed was the implementation, not the thinking. The product judgment stayed the same.

Why this matters for founders

If you're building a product — or hiring someone to build one — the distinction determines what you get.

A vibe-coded product optimises for speed of creation. An agentically engineered product optimises for production quality. Both use AI tools. The difference is the judgment and discipline applied.

The vibe coding reality check shows the consequences of the first approach: 45% security vulnerability rates, 2.74x more issues in AI-generated code, applications that work in demos but fail in production. The AI productivity paradox digs into why AI feels faster while sometimes producing slower results — and how to use it effectively.

The agentic engineering approach is what I've been practicing for over two years — which I call AI-accelerated engineering. Deliberate product decisions, structured specifications (BuildKits), AI execution at speed, human oversight at every step. It's why a 30-day build produces production-ready software rather than a prototype that breaks under real-world conditions.

The skill that matters most

Karpathy specifically noted that agentic engineering "is an art & science and expertise to it" — something people can learn and improve at. The developers who thrive won't be those who prompt fastest. They'll be those who think most clearly about what they're building and why.

For non-technical founders, this means the most important thing isn't learning to code or even learning to prompt effectively. It's having access to someone with the product judgment and engineering discipline to direct AI agents toward production-quality outcomes.

That's the gap the Discovery Sprint fills — bringing 18 years of product experience and 100+ builds of pattern recognition to your specific project.

Frequently asked questions

What is agentic engineering?
Agentic engineering is the practice of directing AI coding agents with structured planning, defined tasks, and rigorous review — treating AI as a capable but supervised implementation team rather than accepting its output without scrutiny. The term was coined by Andrej Karpathy in February 2026 as the evolution of "vibe coding."

How is agentic engineering different from vibe coding?
Vibe coding means prompting AI and accepting the output with minimal review. Agentic engineering means planning the architecture first, breaking work into scoped tasks, having AI agents execute each task, and reviewing every output against quality standards. The METR study showed that unstructured AI use made experienced developers 19% slower — agentic engineering's structure addresses this directly.

Can non-technical founders use agentic engineering?
The strategic principles — plan before building, break work into defined tasks, review output rigorously — apply regardless of technical skill. But the review step requires production experience to evaluate whether AI-generated code is actually correct. For non-technical founders, this means partnering with someone who has that experience rather than relying on AI tools alone.

---

Tom Crossman builds scalable systems and software for service businesses at Hello Crossman. 18 years in product development. Head of Product Engineering at Habito (£3B in mortgages processed). 100+ products shipped. See the case studies →