Yapping is the new coding

In December 2025, Andrej Karpathy crossed a threshold he hadn't anticipated. After watching AI coding agents improve across two years, he began handing over entire chunks of implementation he previously would have written himself. The result was unsettling: for the first time in his career, he felt behind — not because he couldn't produce code, but because the limiting factor had shifted from his typing speed to his capacity to direct work through natural language.

Karpathy later formalized this into two distinct disciplines at Sequoia's AI Ascent 2026. Vibe coding, coined in early 2025, describes the practice of stating what you want and accepting the model's output without rigorous review. It lowers the entry barrier — anyone can prototype an idea. Agentic engineering, the successor concept, treats the model as a capable but fallible collaborator. The developer writes structured specifications, builds evaluation harnesses, and manages multiple agents in parallel while retaining full accountability for quality, security, and system coherence.

This reframing isolates four responsibilities Karpathy argues humans must retain: aesthetics, judgment, taste, and oversight. An agent can implement a feature faster than any single engineer could, but only a human can assess whether the resulting abstraction is durable, whether the approach fits the broader architecture, and whether the output meets production standards rather than just passing visible tests. The model generates; the developer curates.

The counterintuitive outcome is that the quality bar has not dropped — it has risen. Tooling that once screened developers by syntax fluency now filters by reasoning ability. The engineer who thrives is not the fastest typist but the one who can describe intent precisely, evaluate trade-offs across a wide context, and exercise taste without touching the keyboard directly. The ceiling, in Karpathy's framing, extends well beyond what the old 10x engineer label ever captured.