Escrowed Batch Reveal: Eliminating First-Proposal Bias in Agentic Marketplaces Through Visibility Protocol Design

Abstract

\begin{abstract} As LLM-based agents increasingly mediate marketplace transactions, ensuring fair treatment of service providers becomes critical. We identify \textbf{first-proposal bias}: customer agents disproportionately select whichever proposal arrives first, regardless of quality. Under standard protocols where proposals are revealed sequentially, agents select the earliest-arriving option 73.3% of the time---more than double the uniform expectation. We propose \textbf{Escrowed Batch Reveal (EBR)}, a protocol that buffers incoming proposals and reveals them simultaneously in randomized order. EBR reduces first-arrival selection from 73.3% to 24.4% ($p < 0.001$), achieving statistical uniformity across arrival positions. Prompt-based interventions instructing agents to ``wait and compare'' prove insufficient, achieving only 63.3%. Our results demonstrate that fairness in agentic marketplaces requires architectural changes to information flow, not merely behavioral instructions. \textit{WARNING: This paper was generated by an automated research system. The code is publicly available.}\footnote{\url{https://gitlab.com/fars-a/escrowed-batch-reveal-proposal-bias}} \end{abstract}