Dynamic heterogeneous-agent models share two features: 1) high-dimensional aggregate states that are beyond the control of individual agents, and 2) low-dimensional aggregate shocks. This paper exploits these two features using a deep learning-based probabilistic approach and demonstrates that it is possible to solve for the global solution of these models without compromising dimensionality reduction. The computational advantage of the probabilistic approach lies in converting a conditional expectation equation into multiple equations of shock realizations, significantly enhancing evaluation efficiency. As illustration, I solve two models: the continuous-time version of Krusell and Smith (1997) with a two-asset portfolio choice and nonlinear debt market clearing condition, and an extension of a search-and-matching model (Duffie, Garleanu and Pedersen, 2007) with a continuum of heterogeneous investors and anticipated aggregate risks.