Mechanism Design for Multi-Agent Alpha Discovery: Optimizing Agent Distribution in Heterogeneous LLM Markets

We frame multi-agent stock selection as a mechanism design problem: a principal (optimizer) allocates aggregation weight across heterogeneous LLM-powered investor types to maximize portfolio alpha. Drawing on the market disagreement hypothesis, we study mechanisms that reward consensus and penalize disagreement among agents. Through systematic ablation of four mechanism design choices - objective alignment, learnable aggregation weight, optimizer capability, and stability constraint - we demonstrate that these choices interact non-trivially: individual improvements can degrade performance when paired with an insufficient optimizer, and weak stability constraints are worse than none. Critically, at 64 agents the simple baseline outperforms all improved configurations (10d Rank IC $0.033 \pm 0.010$ vs $0.023$ and $0.021$), while a single 512-agent run suggests that the same improvements can become useful at larger scale ($0.044$ vs $0.033$). We also find substantial run-to-run variability that challenges single-seed conclusions common in the literature. These results connect adaptive mechanism design, Stackelberg-style aggregation, and LLM-based strategic reasoning in financial markets.