Earth observations provide an unprecedented view of our changing planet, yet translating these observations into scientific understanding requires new methods that integrate physical models and artificial intelligence. Earth system processes span temporal scales from weather to climate and exhibit nonlinear behavior, including abrupt transitions and climate tipping points, which remain difficult to characterize using conventional modeling approaches alone. We present the AI Climate Tipping-Point Simulator (ACTS), an agentic digital twin framework that combines reduced-order modeling, AI-assisted scientific discovery, forecasting, and AI-driven scientific workflows to accelerate Earth system analysis and prediction. ACTS is enabled by our generalized YETI architecture, which dynamically constructs deep learning models, incorporates foundation models, and orchestrates AI workflows spanning short, medium, and long-range Earth system prediction. Rather than replacing high-fidelity Earth system models, ACTS uses reduced-order models to develop and evaluate AI methods that translate learned representations of Earth system dynamics to more complex Earth system models. Initial accomplishments include the development of reduced-order models for the Amazon and West African Monsoon, AI-enabled discovery of previously unexplored dynamical regimes in ocean circulation models, dynamics-informed forecasting of critical transitions, and a multi-agent framework that coordinates AI and physics-based models to support automated scientific inquiry. Together, these capabilities establish a scalable AI framework for integrating Earth observations, physics-based models, and artificial intelligence to accelerate scientific discovery, advance Earth system prediction across multiple temporal scales, and provide a foundation for future Earth system digital twins.