(Revised August 19, 2026)

Transformers, the core architecture behind modern large language models, combine flexible nonlinear approximation with attention, a mechanism that learns how information is aggregated across inputs. Can attention reveal which parts of an agent distribution matter for aggregate dynamics? I represent cross-sectional distributions as economic tokens and embed a transformer in a recursive Krusell–Smith-style solution method. The transformer forecasts the continuation distribution entering individual Bellman equations. I establish an analytical link between transformer attention and the sensitivity of aggregate policy functions to distributional components, with attention recovering their normalized marginal sensitivities under a restricted representation. In Krusell–Smith benchmarks, the transformer rediscovers approximate aggregation and assigns nearly uniform attention when aggregate capital is nearly sufficient. In a heterogeneous-firm economy with endogenous default, attention concentrates near the firm default margin and becomes state dependent: during downturns attention shifts toward financially fragile firms, whose balance sheets become more consequential for leverage and aggregate dynamics.