A new benchmark called SpatialBench exposes the gap between agent demos and deployable tools by providing 146 verifiable spatial biology analysis problems drawn from real workflows. The suite spans five platforms (Vizgen MERFISH, Takara Seeker, 10x Visium, 10x Xenium, Atlasxomics DBIT‑seq) and seven task categories, emphasizing reproducible outcomes over superficial performance. Authors curated tasks at decision points scientists face—QC, normalization, clustering, cell typing and spatial analysis—so toolmakers can be measured on durability and correctness rather than on curated examples. The benchmark penalizes shortcuts and rewards methods that produce biologically actionable results. SpatialBench signals a move toward rigorous evaluation standards for AI/agent tooling in spatial biology, and it will be used by labs, startups and platform vendors to validate claims and direct engineering toward real‑world utility.