In March 2026, Andrej Karpathy's AutoResearch script ran 50 experiments overnight, revealing the importance of the Karpathy Loop's primitives: editable assets, scalar metrics, and time-boxed cycles. This pattern extends beyond ML training, but teams must also consider the agent's own instruction loop and the need for clear, version-controlled `program.md` documents.
Agent OptimizationAgentsAi Engineering
