Web agents are typically evaluated only on English interfaces, leaving their performance on the world’s multilingual sites largely unstudied. Low‑resource languages (LRLs) suffer especially from a lack of standardized evaluation resources. To bridge this gap, we introduce the Low‑Resource Web Agent Benchmark (LoWAgent), which compares agents across English and four LRLs—Hindi, Bengali, Urdu, and Amharic.\ \ LoWAgent comprises six fully localized websites, each instantiated in five languages, yielding about 3,000 task instances. We evaluate three closed‑source and five open‑source models on all languages and sites. Results show that both closed‑ and open‑source agents degrade on LRLs relative to English, with open‑source agents suffering markedly larger drops.\ \ We also test robustness against realistic interface disruptions such as cookie notices and ad pop‑ups, finding that these further impair agentic performance.\ \ Finally, fine‑tuning on a small subset of LoWAgent traces significantly improves task performance in low‑resource languages, demonstrating the benchmark’s practical utility.\ \ Review: LoWAgent offers a systematic, reproducible platform for multilingual web‑agent evaluation, highlights current models’ weaknesses on LRLs, and shows that modest fine‑tuning can recover performance, providing clear guidance for future research.