Why Traditional Coding Is Becoming Obsolete in the AI Era

Over the last three years, especially the most recent year, the software engineering field has undergone a dramatic shift driven by AI, prompting engineers to reconsider how they keep pace with rapid change.

"I don't think I've typed like a line of code probably since December, which is an extremely large change." — Andrej Karpathy, No Priors podcast, March 2026

Half a Year Without Writing Code

Karpathy, former OpenAI co‑founder and Tesla AI director, disclosed that he has not written a single line of code since December 2025. Instead, he uses natural language to direct AI agents that perform the entire development workflow, turning English into his new programming language.

He distinguishes two concepts: Vibe Coding , which lowers the entry barrier by letting anyone generate runnable applications via prompts, and Agentic Engineering , a disciplined discipline that manages powerful but stochastic AI agents to maintain quality and avoid bugs.

From Software 1.0 to 3.0

Karpathy frames software history in three eras: Software 1.0 (explicit hand‑written code), Software 2.0 (training neural networks on data), and Software 3.0 (prompt‑driven interaction with large language models, which he calls “summoning ghosts”). He notes that LLMs excel in verifiable domains (programming, mathematics) but remain “jagged” in commonsense reasoning.

Verification Gap

Boris Cherny, Anthropic’s engineering lead, observes that AI agents can produce patches faster than humans can verify them, creating a “verification gap.” His remedy emphasizes judgment before delegation, evidence‑based trust, and responsibility after merging code.

Productivity Explosion

Garry Tan compared his own GitHub output in 2013 versus 2026, reporting an 810‑fold increase in logical lines of code, attributing the leap to tooling rather than personal ability.

Engineering as the Final Barrier

When AI removes the scarcity of raw coding speed, the decisive factor becomes engineering rigor: precise specifications, structured skills, and systematic verification. Methods such as the Skills system (prompt‑free, reusable agents), Spec‑Driven Development (contract‑based specifications), and Harness Engineering (building safe runtime environments for agents) are highlighted.

These approaches collectively aim to “structure knowledge to harness unstructured AI capability,” turning fleeting prompts into durable, auditable artifacts.