Uses "Sufficiency Intelligence" to cut AI agent costs while work runs. Sends suitable requests to smaller models, caches repeated reads, trims bulky tool results, keeps the chosen model when unsure, and offers budgets, loop detection and approval controls.