Per-Task Leverage Framework for Human-Agent Collaboration

A recent study published on arXiv introduces a per-task leverage ratio aimed at assessing the effectiveness of human-agent collaboration. This ratio is calculated by taking the amount of human work replaced by an agent and dividing it by the time a human spends defining the task, addressing interruptions during execution, and reviewing outcomes. The denominator is divided into three components, each with unique time-cost factors. The authors demonstrate that information density is directional, with distinct limits for both human-to-agent and agent-to-agent interactions. Asymptotic leverage is analyzed along capability and memory scaling dimensions, with a minimum threshold due to inherent task novelty constrained by human throughput. This framework also applies to a windowed measure for repeated tasks, generated subtasks, and amortized system design costs, with the per-task ceiling not restricting the windowed measure, although both are still limited.