Agent spend is becoming an operational design problem, not just a procurement detail.
In his June 2026 post, Simon Willison discusses Bloomberg’s reporting on Uber’s monthly spending caps for AI coding tools. He sees caps as a sensible response to overspending and questions rewarding token usage for its own sake. That leaves teams with a useful distinction: a spending limit controls exposure; it does not tell you whether the work was worth doing.
The useful question is whether the agent completed work that would otherwise cost more human time, coordination, review, or delay. A run that uses more tokens but produces a reviewed fix may offer better value than a smaller run that leaves someone to diagnose and finish it.
For a hypothetical bug fix, compare a run that needs ten minutes of review and passes its checks with a lower-token attempt that needs two retries and an hour of manual repair. Track the actual model charges, human time, and final result before deciding which approach costs less.
Set a budget and a retry limit for each task, make failed attempts visible, and review outcomes alongside spend. Krosoft builds software that stays useful and gives people time back. Agent workflows deserve the same standard.