Reconstruct Before You Reason: Data Readiness for Multimodal Data Agents

Data for analysis rarely arrives in one place: the values a single computation needs are scattered across relational tables, exports, prose documents, and video. Yet most work on data agents concentrates on the analyst, the reasoning that turns a question into a query, and evaluates it on data already curated into a clean, queryable form. The KDD Cup 2026 DataSpace benchmark drops that assumption: each task hands the agent a bundle of heterogeneous sources that jointly encode one latent relational dataset and demands an exact answer, with the language model fixed, so what distinguishes two systems is how each prepares and exposes evidence rather than how strong its model is. This design lets us treat readiness and reasoning as separate questions. In the competition's structured-only first phase our multi-agent text-to-SQL analyst placed in the top ${\sim}$1\% of 703 registered teams; we take that analyst as a fixed precondition, and the second phase adds narrative documents and video, letting us study data readiness on top of it. Holding the model fixed and materializing each bundle into a canonical, queryable database, we intervene on one component at a time and find that readiness, not query-time context, is the bottleneck: the load-bearing reconstruction steps each cost 0.13-0.31, whereas removing any single query-time component stays within a 0.05 noise band, and removing all of them costs only 0.08. Full reconstruction reaches 0.81 $\pm$ 0.05, and even a plain text-to-SQL analyst over the reconstructed tables reaches 0.73. On the blind evaluation, the complete reconstruction-first system finished in the top ${\sim}$2\% of 703 teams, consistent with the design generalizing beyond the public set. For agents that analyze scattered data, reliable analysis begins with reconstructing the analysis-ready dataset a question requires, not with better reasoning over raw evidence.