Ten Queries Beat Five Hundred: Closed-Loop Agentic Search for Transportable Sleep-EEG Interfaces

We study downstream interface selection under institutional shift using a sleep-EEG foundation model whose 539 physiological heads, organized into 11 biological categories, define more than 3.4 billion possible four-head interfaces. A generalist controller plans experiments from source-institution feedback and never receives held-site labels or scores. After ten evaluations, the agentic search reaches a macro held-site age-conditioned AUROC of 0.630, compared with 0.564 for matched random search and 0.550 for evolutionary search. The resulting selections also exceed the 500-candidate random-bank point estimate with one-fiftieth as many evaluations. At I0006, which contributes 441 eligible age-matched pairs, the margins are $+0.114$ $[+0.004,\,+0.224]$ over matched random search and $+0.142$ $[+0.004,\,+0.275]$ over evolutionary search. Within the queried sets, the average gap between the selected interface and the best held-site candidate is 0.100, compared with 0.239 for the 500-candidate bank. Even the weakest agent-guided run exceeds the median run of both broad-search baselines. Although taxonomy labels are available in the head catalog, the controller is never instructed to diversify across them; its selected quartets nevertheless span 3.3 to 3.4 physiological categories. Separate runs show mean pairwise head overlap of 0.203--0.259, compared with 0.005--0.013 for evolutionary and 0.048--0.075 for taxonomy-diverse search. An independently developed, substantially smaller controller reaches 0.600 at $B{=}10$ ($[0.514,\,0.679]$) and finishes within 0.007 of the expert prior. Across both controllers, the strongest searches are compact, adaptive, and biologically structured.