A technical deep dive into Karpathy’s thesis: agents are promising but immature. The bottlenecks are memory, multimodality, reliable computer use, better learning signals, and better self-improvement loops.