Researchers have built a test to measure how badly AI agents fall apart when they have to work across multiple languages at once.
PolyWorkBench is a new benchmark covering 67 tasks spread across five workplace domains — commerce, knowledge work, legal analysis, localization, and manufacturing. Unlike most existing benchmarks, which quietly assume everything happens in one language, PolyWorkBench throws agents into workflows where inputs, reasoning steps, tool calls, and outputs may all involve different languages simultaneously. The evaluation framework combines structural grading, executable verification, and model-based semantic scoring to check both whether the agent got the job done and whether its outputs were linguistically coherent.
The results are a useful corrective to the hype around agentic AI. State-of-the-art LLM agents show significant performance drops in multilingual settings compared to single-language equivalents — and the degradation compounds across steps, meaning a small stumble early in a multi-step workflow tends to snowball. That matters because real enterprise deployments — the ones AI labs are pitching hardest right now — are almost never monolingual.
Most agent benchmarks have been built around English workflows, which flatters English-dominant models and papers over a gap that will matter the moment these systems touch a global supply chain, a cross-border legal document, or a customer base that does not default to English. PolyWorkBench does not tell us how to fix the problem, but it at least names it.