Preprint reports lower latency and token use in multi-agent workflow tests

A newly posted preprint describes token and context-management patterns for multi-agent AI workflows, drawing on an internal production dashboard that routes LLM-generated work summaries. The authors report cold-load latency of 61 to 116 seconds across six timed runs, versus an operational baseline of roughly 3.5 to 10.5 minutes, and estimate a 60% to 70% token reduction. The paper also reports controlled context-composition tests across 11 model configurations.