World models let agents anticipate the consequences of their actions, but enterprise systems pose a twist: the rules are tenant-specific, change over time, and are often readable at runtime. So does an agent really need to internalize dynamics it could simply look up? Agents that discover transition logic at inference time, by reading the system's live configuration, stay robust under deployment shift where offline-trained world models degrade. Backed by CascadeBench, a new benchmark for enterprise cascade prediction, these results suggest a rethink of when learned world models are worth the trouble. Presented by Jishnu S Nair, on October 2, 2026. Paper: https://arxiv.org/pdf/2605.12178 Dataset: https://huggingface.co/datasets/ServiceNow-AI/cascade\_bench More from ServiceNow AI Research: https://www.servicenow.com/research/