EarthRecord is a benchmark built on raw satellite archives, spanning 56 sites and four decades of Landsat, MODIS, VIIRS, and Sentinel‑2 observations. Unlike existing benchmarks, it does not supply curated images nor restrict sensor usage; cloud masking, cross‑sensor harmonization, and other preprocessing steps are left to the multimodal agents themselves. The benchmark poses 238 land‑change questions across ten analysis tasks—such as land‑cover classification, change detection, and trend analysis—and scores answers against hidden references from global products, agreement among methods, and visual interpretation.\n\nTwo control experiments probe the source of the answers:\n- Closed‑book: the entire historical record is removed, leaving only current imagery;\n- Instrument‑pair: data from one satellite are deliberately altered while the ground truth remains unchanged.\n\nEvaluations of twelve leading multimodal agents show that while most agents can complete end‑to‑end analyses, they struggle when a response requires mapping directly from imagery. Further analysis reveals that the record shifts answers from prior knowledge to the pixel level, yet it does not yet capture real ground change: when a single satellite’s data are altered, about half of the correct answers follow the alteration.\n\nEarthRecord thus offers a principled platform for measuring and advancing agents from raw pixels to reliable land‑change insights.\n\nReview: The benchmark exposes current multimodal agents’ limitations in handling raw remote‑sensing data at scale and points to concrete research directions for improving real‑world change detection.