Submitted to arXiv on September 1, 2026 (paper 2609.01056) by Leonardo Ranaldi, Sherrie Shen, Jushi Kai and Alexandra Birch. Their argument is that agent benchmarks have converged on English, single-outcome tasks and therefore miss three things that matter in deployment: whether the agent preserves the state of the environment it acts in, whether it performs across languages, and whether the scenarios resemble the grounded everyday workflows real users bring.
WorldBench is a multilingual benchmark of persona-grounded everyday workflows in which agents act through structured actions inside a sandbox. It comprises 1,600 tasks spanning seven languages and eight cultures, filtered and refined with feedback from human annotators holding language- and culture-specific expertise rather than translated from an English seed set. Evaluation extends prior metrics and introduces Constrained Task Success, or CTS, which combines natural language instructions with testbeds to score task completion, minimal modification and complementary properties, using a mixture of deterministic checks and LLM-as-a-Judge.
The result is that frontier models reach only 49.2 percent CTS, and every model shows a large gap between correctness and environment preservation. In plain terms, the agents frequently get the answer right while leaving collateral damage behind - changing state they were not asked to change. That gap widens on long-horizon tasks and under explicit state-preservation constraints.
For anyone deploying agents against real systems, the environment-preservation half of that score is the part with the operational teeth. A benchmark that only asks “did the task succeed” cannot distinguish an agent that did the job from one that did the job and also modified four unrelated records; WorldBench is built to separate the two, and finds that current models routinely fail the second test in exactly the multilingual, long-horizon settings a global rollout would encounter first.