Quote-Batched Payment Protocol for Reducing First-Proposal Bias in Agentic Marketplaces

Abstract

As large language models are increasingly deployed as autonomous agents in economic transactions, their systematic biases can distort market outcomes. We study first-proposal bias---the tendency of LLM customer agents to disproportionately select the first proposal they receive---in agentic marketplaces. We propose QuoteBatch, a mechanism design intervention that combines a hard-gate blocking payment until multiple proposals arrive with anti-anchoring prompt instructions. On Claude claude-sonnet-4-5, QuoteBatch reduces first-proposal bias from 100% to 6.7% (93.3 percentage point reduction, $p<0.001$) while maintaining 100% task completion. However, the same intervention yields only a 10 percentage point reduction on Gemini gemini-2.5-flash (not statistically significant), revealing substantial heterogeneity in how different LLMs respond to mechanism design. Our findings highlight that deploying AI agents in economic systems requires model-specific bias mitigation strategies.