"Automate only what can be objectively evaluated Karpathy cautions that autonomous AI loops only work reliably when success can be measured without human judgment: "this is extremely well suited to anything that has objective metrics that are easy to evaluate" — he uses writing CUDA kernels as the ideal example, where you can verify the output is both correct and faster. Conversely, anything "softer" falls outside the reinforcement-learning rails and the system meanders. Before delegating a task to an autonomous agent, ask yourself: can I define a clear, measurable criterion for success? If not, keep humans in the loop for that step. Source: Andrej Karpathy, Conviction, Episode with Andrej Karpathy"