Local inference finally helps agent loops when latency stays under a second. The win is not raw scores. It is the feedback cycle when you can afford to retry without burning API budget.