Title: Latent Intent Mapping in Heterogeneous Agent Collectives via Gradient-Based Future Surrogates Abstract: This research introduces a framework for extracting and representing latent intent within cooperative, heterogeneous multi-agent systems governed by deep reinforcement learning policies. We operationalize intent as a predictive function of future environmental states and critical decision points, inferred not from world models but from internal policy gradients and temporal agent memory. By treating feature attribution maps—specifically Integrated Gradients—as auxiliary signals, we train surrogate neural networks to forecast agent positions and identify high-certainty states before they occur. Experiments in mixed-composition environments (Simple Spread and Knights-Archers-Zombies) reveal that gradient-based explanations significantly enhance event prediction, while methods like Shapley values offer negligible gain. We further examine the utility of recurrent cell states and stacked observations as intent proxies, finding that memory-enhanced policies yield more predictable future traces. The study concludes that in heterogeneous multi-agent settings, intent can be partially decoded via differential explanations of policy logits, offering a post-hoc, model-agnostic path toward interpretable coordination. This approach bridges explainable AI with multi-agent behavior analysis, providing a diagnostic tool for verifying cooperative adherence and forecasting emergent group dynamics.