From Vibe Coding to Agentic Engineering: The Discipline That Will Define the Next Decade of Software

In early 2025, Andrej Karpathy introduced the term “vibe coding” to describe a new way of building software — describing what you want and letting AI generate the rest. It felt liberating. For many, it still does.

But in 2026, Karpathy has moved on. In a recent conversation at Sequoia’s AI Ascent, he made a sharp distinction: vibe coding raises the floor, while agentic engineering raises the ceiling.

Vibe coding is fast, accessible, and often surprisingly effective for prototypes and personal tools. It democratizes creation. Yet it also carries a hidden risk: the illusion of progress without the discipline of engineering.

Agentic engineering is something else entirely.

It treats AI agents not as magical code generators, but as jagged, fallible, stochastic collaborators — powerful in narrow domains, yet prone to subtle errors, hallucinations, and compounding mistakes. The role of the human shifts from typing code to orchestrating these agents with precision.

This new discipline demands new practices:

• Clear specifications instead of vague prompts
• Structured decomposition of work
• Parallel agent teams with defined responsibilities
• Rigorous verification loops and evaluation metrics
• Explicit oversight, permission boundaries, and rollback mechanisms
• Preservation of taste, security, and long-term maintainability

Karpathy’s message is unambiguous: the era of casual prompting is giving way to a more serious craft. One that resembles traditional engineering more than it does magic.

For organizations, this distinction matters profoundly. Vibe coding can accelerate experiments. Agentic engineering is what separates promising prototypes from production systems that actually deliver lasting value.

The organizations that will win are not those that adopt AI the fastest, but those that learn to direct it with discipline.

Technology is easy. Directing it with judgment is hard.