AI agents using large language models (LLMs) automate tasks in software. Testing many sequences of actions, a technique called , can improve them, but it is risky in shared systems because actions can cause irreversible changes or interfere with each other.
Agentic World Modeling avoids this by simulating actions through safe read-only queries to the system, then reasoning over the responses to predict outcomes without altering the shared environment.
Across three benchmarks, it outperformed a simple simulation baseline by 40.07% and a trained world model (a model that predicts environment changes) by 12.36%. This makes test-time scaling safe for deploying agents in realistic settings.