OpenAI cofounder Andrej Karpathy says it will take a decade before AI agents actually work (www.businessinsider.com)

OpenAI cofounder Andrej Karpathy told the Dwarkesh Podcast that functional, autonomous AI agents are still far from reality and "will take about a decade" to mature. He argued current agents "just don't work": they lack multimodal competence, reliable computer use, continual learning and durable memory, and generally fall short of the cognitive abilities needed to autonomously complete complex tasks. Karpathy doubled down on X, criticizing industry hype that oversells tooling and a future where humans are made "useless" by fully autonomous systems — a future he doesn't want, especially if it produces ubiquitous low‑quality "AI slop." Technically, Karpathy's critique highlights core gaps: grounding (accurate API usage and docs verification), interactive clarification (asking when uncertain), persistent memory and continual learning, and robust multimodal perception. He and others point to error compounding as a practical limiter — if each action has ~20% error, multi‑step tasks rapidly degrade reliability (five steps yields ~32% chance of full success). The implication for researchers and product teams is clear: prioritize trustworthiness, evaluation metrics for multi‑step reliability, better grounding and memory systems, and human–AI collaboration workflows rather than betting on fully autonomous agents by 2025.