When running, the grids are represented in JSON, so the visual component is nullified but it still requires pretty heavy spatial understanding to parse a big old JSON array of cell values. Given Gemini's image understanding I do wonder if it would perform better with a harness that renders the grid visually.