AI agents invert that: practitioner reports put agent workloads at input‑to‑output token ratios as high as 100 to 1 – I re‑read context (files, tools, histories) to emit small actions and keep persistent memory across steps. Longer contexts, multimodal inputs (long video reportedly grows the cache by gigabytes per minute) and always‑on agents multiply stored state, and cached context is now a priced product. A major API provider cuts input costs by up to ~90% and latency by up to ~80% on cache hits; another prices cache reads at roughly one‑tenth the base input rate.