Existing scientific‑agent benchmarks rarely test whether models can faithfully reconstruct the exact molecular structures and generation workflows described in sequencing protocols, a capability essential for analyzing custom sequencing assays. To address this gap we introduce LibStructBench, a document‑grounded benchmark that evaluates agents on reconstructing sequencing‑library structures and molecular workflows.\ \ Our experiments show that state‑of‑the‑art agents often produce approximately correct reconstructions, yet they remain unreliable for molecular‑state representation and workflow connectivity, failing to achieve exact replication. Switching the underlying language model yields larger average performance differences than swapping agent harnesses, with a statistically significant effect only for molecular‑state reconstruction.\ \ In a small pilot, adding extra domain knowledge improved oligo‑family recall but did not raise the mean molecular‑workflow scores. Autonomous accumulation of procedural rules and cross‑protocol examples exhibited non‑monotonic validation performance and no mean held‑out molecular‑workflow improvement along the tested learning trajectory.\ \ These findings highlight the need for more controlled and verifiable approaches to scientific reasoning about molecular processes in complex experimental protocols.\ \ Review