Andrej Karpathy used an appearance on the Dwarkesh Podcast to pour cold water on the agent boom, saying today’s agents “just don’t work” and that it will take roughly a decade to fix what ails them. “They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use,” he said. “They don’t have continual learning. You can’t just tell them something and they’ll remember it. They’re cognitively lacking and it’s just not working.”

Karpathy says the bottleneck is making agents actually work. He says to think of them like an employee or intern you would hire to work with you, and says the early agents he uses daily, like Claude and Codex, are impressive but still far from that standard. He says the problems are tractable, surmountable, but still difficult.

He was also careful to distance himself from outright skeptics, estimating his timelines are five to ten times more pessimistic than the typical San Francisco AI house party, but still optimistic about the field overall.