AI coding agents reshape software development, Andrej Karpathy notes shift

AI coding agents have rapidly changed software development practices, moving many tasks from manual coding to automated, multi‑agent workflows. Andrej Karpathy, a prominent AI voice, posted that he feels “behind as a programmer” and described a new programmable abstraction layer involving agents, prompts, memory, tools and plugins. He highlighted a shift in his own workflow from 80% manual coding to 80% agent‑driven coding within weeks, noting that he now programs largely in natural language.

Karpathy also shared observations from extensive use of Anthropic’s Claude Code, pointing out agent pitfalls such as wrong assumptions, lack of clarification, and overly sycophantic behavior. He warned that agents can generate bloated code that must be trimmed, and cautioned developers to monitor them closely. Maximilian Schwarzmüller provided a contrasting, reassuring video message, suggesting a balanced view of the fast‑moving AI coding landscape.