From Vibe Coding to Agentic Engineering: What Karpathy's New Term Means for Builders

The Guy Who Named It Says It's Already Evolving

In February 2025, Andrej Karpathy posted on X: "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."

That phrase rewired how an entire industry thought about AI-assisted development. We wrote our own Vibe Coding Manifesto about it. The term showed up in MIT Technology Review's 2026 Breakthrough Technologies list. It became shorthand for a genuine shift in who could build software and how.

On February 4, 2026, Karpathy posted again. This time with a course correction.

"Programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny." He said many people had tried to name what comes after vibe coding, and his favorite was "agentic engineering."

His reasoning: "'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."

This isn't just a rebrand. It's an acknowledgment that vibe coding hit a wall, and the industry needs something more rigorous to get past it.

What Changed in 12 Months

When vibe coding took off in early 2025, the appeal was obvious. Describe what you want. AI writes it. Ship it. Don't even read the code if you don't want to. For prototypes, personal projects, and quick experiments, it was revolutionary.

Then people started building real things with it.

The quality problems surfaced fast. GitClear's analysis of 211 million changed lines of code found the share of refactored lines fell from 25% in 2021 to under 10% in 2024, while copy-pasted lines rose from 8.3% to 12.3%. A December 2025 CodeRabbit study of 470 pull requests found AI-co-authored PRs had about 1.7x more issues overall, and security issues up to 2.74x higher in the worst category (cross-site scripting).

Teams lost understanding of their own systems. Software researcher Margaret-Anne Storey wrote about "cognitive debt," a term gaining traction for the growing gap between what AI builds and what teams actually understand about their systems. Simon Willison, one of the most respected voices in the developer community, admitted he'd found himself "getting lost in my own projects."

Open source maintainers revolted. cURL ended its bug bounty in January 2026, after its maintainer had put AI slop at about 20% of all submissions in 2025. Ghostty closes drive-by AI pull requests and bans people who submit bad AI-generated code. tldraw began automatically closing pull requests from external contributors. RedMonk's Kate Holterhoff called it the "AI Slopageddon".

Vibe coding works for getting started. It doesn't work for building things that last. Karpathy saw that, and he picked a name for the next phase.

Vibe Coding vs. Agentic Engineering

The difference isn't subtle.

Vibe coding says: prompt the AI, accept what comes back, ship it. The developer is along for the ride. Understanding the code is optional. The vibe is the product.

Agentic engineering says: orchestrate AI agents to plan, write, test, and ship code under structured human oversight. The developer is the architect. Understanding is mandatory. The engineering is the product.

In practice, agentic engineering looks like this:

You design the architecture first. Before any AI writes a line of code, you define the system structure, the interfaces, and the constraints.
You break work into well-defined tasks that agents can execute independently.
You run multiple agents in parallel, each working on isolated pieces of the system.
You review everything. Not just "does it compile?" but "does this fit the broader system design and does anyone on the team understand why?"

Anthropic's 2026 Agentic Coding Trends Report cites its own research finding that developers use AI in roughly 60% of their work but can "fully delegate" only 0 to 20% of tasks. That's the ratio. AI is in most of the work. Humans still supervise nearly all of it.

The Tools Have Caught Up

Part of why agentic engineering is viable now is the tooling matured dramatically in the past year.

Claude Code turned AI-assisted development into a terminal-first workflow where you can run multiple agent sessions in parallel on complex multi-file refactors. OpenAI Codex launched a desktop app for running multiple coding agents in parallel on the same repo. Cursor says more than 30,000 NVIDIA developers use it daily. GitHub Copilot added autonomous agent mode. Devin and Windsurf pushed fully autonomous coding further.

But the tool that matters most is the one between your ears.

MIT Technology Review named "generative coding" a 2026 breakthrough technology, noting that AI now writes as much as 30% of Microsoft's code and more than a quarter of Google's, according to the heads of those companies. Mark Zuckerberg wants most of Meta's code written by AI agents in the near future. The scale is real.

But the same article noted the core challenge, citing MIT CSAIL researchers: AI-generated code that looks plausible may not always do what it's designed to do. Tools generate code. Engineers verify it works.

What This Means If You're Not a Developer

I know a lot of our readers aren't writing code themselves. Here's why agentic engineering still matters to you.

The barrier between "people who build software" and "everyone else" is dissolving. Domain experts in marketing, operations, finance, and design are starting to build their own tools and automations using AI agents. You don't need to understand Python. You need to understand your problem well enough to direct an agent.

But directing an agent is a skill. This is the "engineering" part of agentic engineering. It's not just prompting. It's knowing how to break a problem into pieces, define what "done" looks like, and evaluate whether the output actually solves the problem. Those are transferable skills from any professional domain.

The cost of building just dropped again. Work that took weeks can now happen in days. Per Anthropic's report, one Augment Code enterprise customer finished a project its CTO had estimated at 4 to 8 months in two weeks. Zapier reports 89% AI adoption across the whole company, with more than 800 agents deployed internally. At Rakuten, Claude Code implemented a complex method in vLLM, a 12.5 million line codebase, in seven hours of autonomous work, matching the reference method with 99.9% numerical accuracy.

MarketsandMarkets projects the AI agents market will grow from $7.84 billion in 2025 to $52.62 billion by 2030. This isn't a fad. It's infrastructure.

Where We're Headed

The trajectory from the last 12 months tells a clear story.

Phase 1 (2025): Vibe coding. "Just let the AI write it." Revolutionary for accessibility. Terrible for sustainability.

Phase 2 (2026): Agentic engineering. "Orchestrate AI agents with human oversight." Better quality. Still requires judgment and expertise.

Phase 3 (ahead): We'll see. But the pattern is clear: every wave of AI-assisted development gets more powerful and more structured. The "just prompt and pray" era lasted about a year before the industry demanded rigor.

If you're building with AI right now, whether it's software, content, or business processes, the lesson from vibe coding's evolution applies universally. Speed without structure creates debt. Speed with structure creates value.

The vibes got us started. The engineering gets us where we're going.

Want to bring agentic engineering into your business workflow? Book a free 30-minute call and let's figure out where AI agents can do the heavy lifting while your team stays in the driver's seat.

Sources: