The Automated but Risky Game

Shenzhe Zhu and five coauthors study delegated consumer negotiations between language-model agents. This entry covers arXiv v4, whose expanded title includes “Modeling and Benchmarking.”^[1]

Method and contribution

Buyer and seller agents negotiate over 100 products spanning electronics, vehicles, and property. Budgets, costs, roles, and stopping rules define the simulation. Models differ in negotiated outcomes and sometimes violate transaction constraints.^[1]

Alignment relevance and limits

The findings connect Agentic Inequality with Multi-Agent Risks: a useful assistant may still disadvantage its user or accept an invalid deal. These are simulated transactions, not observed consumer losses. Wholesale costs are estimated by GPT-4o rather than independently established; a model judge also classifies negotiation decisions. Prompt design and the gap between simulation and actual markets limit generalization. Differences between models do not isolate a single causal effect of intelligence. Compare The Illusion of Rationality: Tacit Bias and Strategic Dominance in Frontier LLM Negotiation Games for a separate bargaining evaluation.