The performance of multi‑agent LLM systems depends not only on the capabilities of the underlying models but also on information‑sharing policies that determine what each agent observes and at what granularity. To test whether more detailed information always improves collective outcomes or can create a transparency paradox, we introduce an empirical information design framework that systematically compares disclosure policies across economic markets while keeping interaction rules and models fixed. Experiments cover multi‑resource trade, bilateral bargaining, and double auctions, evaluating ten frontier LLMs (including GPT‑6 Sol, Claude Sonnet 5, Claude Opus 5.5, Grok 4.7, Gemini 3.1 Pro) and using ground‑truth utilities to separate identified opportunities from realized results. Finer disclosure yields highly model‑dependent effects. In trade, revealing exact valuations instead of ordinal preferences reduces GPT‑6 Sol’s realized gains from 98% to 91% of the maximum, yet boosts Claude Sonnet 5 from 72% to 86%. In bargaining, exact information leads Grok 4.7 to produce less balanced agreements, whereas Claude Opus 5.5 splits gains symmetrically. In double auctions, full disclosure inflates price demands, causing market breakdown and welfare drops for Gemini 3.1 Pro (97% → 78%), Claude Opus 5.5 (94% → 64%), and Grok 4.7 (97% → 62%). Overall, the findings suggest that model selection and information design must be co‑optimized, as LLMs vary markedly in how their strategic responses to information affect system objectives.
Review: This work highlights that in multi‑agent LLM economies, transparency is not universally beneficial; careful alignment of information policies with specific model behaviors is essential for maximizing overall efficiency.