ROUTE8: An Agent for Precise Information Extraction from the Biological Literature

Literature research in biology often comes down to precise values in supplemen- tary tables, figure panels and database releases. We present ROUTE8, an agent that helps researchers find such information across the papers on a biological topic and extract it accurately. It is deliberately simple: one tool-use loop with web search, an image viewer, and a bash tool in a bubblewrap sandbox, where the model downloads papers, supplements and data files and computes over them with code. To measure whether it meets this aim, we use the seven no-context retrieval sub- sets of LABBench2 (821 tasks). Their answer keys are public, so for evaluation the sandbox’s only network path is a logging, egress-filtering proxy. With Claude Opus 5 as the backbone, the official harness and the official grader, ROUTE8 scores 78.7% micro accuracy (mean of two runs) under a conservative scoring rule, and 80.5% under the official rule. The best published official-protocol result is 63.9% (GPT-5.2-pro); paired per task against it, ROUTE8 is ahead by 14.4– 15.2 pp. It is significantly ahead on SuppQA2, DbQA2, TableQA2 and FigQA2, statistically tied on LitQA3 and TrialQA, and significantly behind on PatentQA. Against the same, newer backbone with its vendor’s built-in tools (degraded: the vendor code-execution tool was unavailable to us), ROUTE8 is ahead by 20.6 pp micro. Ablations trace most of the gain on database and supplement tasks to the shell’s own network access; removing the prompt’s cheat-sheets has no significant effect. We also describe a contamination and inflation audit for web agents on benchmarks with public answer keys. It found two integrity failures in our own pre-audit agent, both fixed before the reported runs, and no evidence of answer- key access in the reported runs.