Overview

• OpenAI cofounder Andrej Karpathy argues current autonomous agents are not usable due to limited intelligence, lack of multimodality, unreliable computer operation, and weak persistent memory or learning.
• He estimates it will take about ten years to address these core deficits and reach broadly useful, reliable performance.
• He criticizes the industry for releasing agent tools faster than underlying capabilities justify, warning this accelerates low‑quality outputs or so‑called AI slop.
• He cautions that aiming to replace workers with agents could devalue human contributions and urges designs that augment people instead.
• He describes preferred behaviors for collaborative systems, including auto‑surfacing API documentation, showing how tools are called, and asking clarifying questions rather than guessing.