Evaluating and improving AI agents for sequencing protocol understanding

Existing scientific-agent benchmarks do not test whether agents can reconstruct the exact molecular structure and workflow of sequencing protocols, which is critical for analyzing customized sequencing assays. We introduce LibStructBench, a document-grounded benchmark for reconstructing the molecular workflow of sequencing-library generation. Our experiments show that frontier agents are often approximately correct but remain unreliable at exact reconstruction, with frequent discrepancies in molecular-state representation and workflow connectivity. Switching the underlying LLM produced larger average performance differences than switching agent harnesses, with a statistically significant difference only for molecular-state reconstruction. Additional knowledge improved oligo recovery but did not improve molecular-workflow reconstruction. Autonomous accumulation of procedural rules and cross-protocol examples likewise failed to improve held-out molecular-workflow reconstruction in the tested learning trajectory. These findings highlight the need for more controlled and verifiable methods for scientific reasoning over molecular processes in complex experimental protocols.