The paper presents a machine learning framework that integrates manifold learning, neural networks, Gaussian processes, and Equation-Free (EF) multiscale approaches to construct effective reduced-order models (ROMs) from detailed agent-based simulators. The framework aims to detect tipping points and quantify the uncertainty of rare events near these points. Two illustrative examples are used: an event-driven, stochastic financial market model and a compartmental stochastic epidemic model on an Erdős-Rényi network. The framework reveals that the emergent dynamics around tipping points can be effectively described by a one-dimensional stochastic differential equation, reducing computational costs significantly. The methods include discovering low-dimensional latent spaces using Diffusion Maps, identifying mesoscopic Integro Partial Differential Equations (IPDEs) and mean-field Stochastic Differential Equations (SDEs) via machine learning, and performing bifurcation analysis and rare-event analysis using the identified models. The results show that the proposed framework accurately locates tipping points and provides efficient computational benefits compared to full agent-based simulations.The paper presents a machine learning framework that integrates manifold learning, neural networks, Gaussian processes, and Equation-Free (EF) multiscale approaches to construct effective reduced-order models (ROMs) from detailed agent-based simulators. The framework aims to detect tipping points and quantify the uncertainty of rare events near these points. Two illustrative examples are used: an event-driven, stochastic financial market model and a compartmental stochastic epidemic model on an Erdős-Rényi network. The framework reveals that the emergent dynamics around tipping points can be effectively described by a one-dimensional stochastic differential equation, reducing computational costs significantly. The methods include discovering low-dimensional latent spaces using Diffusion Maps, identifying mesoscopic Integro Partial Differential Equations (IPDEs) and mean-field Stochastic Differential Equations (SDEs) via machine learning, and performing bifurcation analysis and rare-event analysis using the identified models. The results show that the proposed framework accurately locates tipping points and provides efficient computational benefits compared to full agent-based simulations.