Can Self-Improving Agents Recover Cofolding Failures?

Co-folding models predict protein--ligand complexes directly from sequence and ligand chemistry, but no single model dominates, and each ranks its own samples with a confidence score calibrated only against itself. On our held-out set, a near-native pose (ligand RMSD $<2$\,\AA) exists among 20 candidates from four engines for 87.9\% of ligands, whereas the best engine's own Top-1 reaches 73.3\%: recovery is largely a selection problem. We introduce FoldScope, a framework that tests whether an LLM agent can perform this selection \emph{without} native confidence scores and improve its own procedure from experience. A Solver agent opens anonymized engine groups one at a time, under a mean cap of three of four groups, and selects one complete complex from reference-free structural evidence. A separate Improver agent reads Solver traces and reference-based feedback on one half of the training data and rewrites the procedure itself, its instructions and analysis code; a revision is adopted only if it outperforms the current procedure on the other half. With Claude Opus~5.5, five rounds of self-improvement raise test Top-1 accuracy from 69.9\% to 73.5\% and mean lDDT-PLI from 0.817 to 0.830, slightly exceeding Boltz-2's native ranking (73.3\%, 0.822) while outperforming other models.