Dataset - Evaluating Single-Agent and Multi-Agent AI Architectures for Analytical Modeling

This is the data required for "Evaluating Single-Agent and Multi-Agent AI Architectures for Analytical Modeling" paper. Abstract As Artificial Intelligence (AI) Agents are adopted in analytical workflows, a key design question arises: can a single agent manage modeling complexity, or do multi-agent frameworks justify their added cost? We empirically evaluate a single-agent architecture that integrates exploration, modeling, validation, and revision with a multi-agent architecture that separates analyst, critic, and refiner roles across descriptive and predictive tasks. For well-defined descriptive tasks such as summary statistics and deterministic transformations, the single-agent approach achieves quality comparable to multi-agent designs while requiring fewer model invocations and lower token usage, making coordination overhead unnecessary. In contrast, for predictive tasks involving ambiguity and nonlinear feature relationships, multi-agent workflows more reliably preserve true predictive signals that may be missed by shallow linear screening, though at higher computational cost. Overall, these results indicate that multi-agent architectures enhance reliability under uncertainty, whereas single-agent designs are appropriate for routine, low-risk tasks.