token budget is becoming part of my agent workflow design

About this article

I think token budget is becoming part of agent workflow design. If every run feels expensive, people under-test. They save quota, overthink prompts, and avoid the repetition that reveals failure modes. If every run feels cheap, people can over-delegate. They generate more output than they can review. So the useful question is not "which model is best?" It is: Which step deserves which level of model? My current rule: cheap / lower-reasoning runs for bounded, reviewable repetition stronger mod...