Test-time scaling (TTS) boosts large language model agents by sampling and selecting among multiple candidate trajectories. Existing implementations assume these trajectories can be run in parallel, an assumption that fails in real deployments where enterprise workspaces and cloud services are shared, persistent, costly to replicate, and actions may cause irreversible side effects or cross‑trajectory interference. Current generative world models compress the environment state into a single prompt, limiting reliable simulation under partial observability.\
To address these issues we introduce Agentic World Modeling (AWM). AWM reformulates next‑observation prediction as a multi‑turn evidence‑acquisition task: for each hypothetical action, an agentic simulator adaptively queries the real environment through safe read‑only interfaces, reasons over the returned evidence, and uses executable computation to construct the resulting observation without mutating shared state.\
When integrated with trajectory‑level TTS, AWM enables parallel exploration over a single shared environment. Experiments on $\tau$‑bench, AppWorld, and OfficeBench show that AWM outperforms a single‑shot simulation baseline by 40.07% on average and surpasses a trained world‑model baseline by 12.36% on average. These results demonstrate that interactive, read‑grounded simulation provides a practical foundation for safe and scalable agentic test‑time scaling in non‑resettable shared environments.\
Review: AWM’s read‑only interaction and executable reasoning break the reliance on environment replicas, offering a viable path for deploying large‑model agents safely in real systems.