Authors introduced SpatialBench, a suite of 146 verifiable spatial‑biology analysis problems across five platforms and seven task categories, designed to benchmark AI agents on real laboratory workflows. The paper argues current agent demos look convincing but fail in realistic, messy datasets; SpatialBench supplies reproducible tasks (QC, clustering, cell typing, spatial analysis) to measure true utility. The benchmark tests agents’ ability to produce verifiable biological results rather than polished narratives, aiming to accelerate reliable deployment of AI in spatial biology.