Agent-based simulations are indisputably effective for analyzing complex processes such as traffic patterns and social systems. However, human experts often face the challenges in repeating the simulation many times when evaluating a large variety of scenarios. To reduce the computational burden, we propose an approach for inferring the end results in the middle of simulations. For each simulated scenario, we design a feature that compactly aggregates the agents’ states over time. Given a sufficient number of such features we show how to accurately predict the end results without fully performing the simulations. Our experiments with traffic simulations confirmed that our approach achieved better accuracies than existing simulation metamodeling approaches that only use the inputs and outputs of the simulations. Our results imply that one can quickly evaluate all scenarios by performing full simulations on only a fraction of them, and partial simulations on the rest.