Open-Rosalind: Tool-First Biomedical LLM Agents with Process-Aware Benchmarking

Large language models are increasingly used as scientific agents, yet the flexibility that benefits general-purpose agents can conflict with the accountability required in biomedical research. We study whether biomedical agents can be organized around auditable constraints rather than unconstrained autonomy. We present Open-Rosalind, a tool-first bio-agent system designed around four operational principles: evidence-grounded outputs, trace completeness, workflow-constrained execution, and explicit tool mediation for factual claims. To evaluate these principles, we introduce Open-Rosalind BioBench, a process-aware benchmark that measures not only task accuracy but also tool correctness, citation presence, trace completeness, and failure rate. On a strict in-house benchmark, the reference pipeline achieves 81.4% accuracy with complete execution traces. In multi-model ablations and paired replications, removing tools reduces accuracy by 19.3 to 26.4 percentage points, indicating that tool-first execution is the strongest and most stable contributor to performance. Constrained workflows also reduce lower-tail failures for models that are weak at free-form tool use. However, an author-independent 30-task hold-out initially revealed severe external-validity collapse on the deployment model. After diagnosing five routing and normalization failures and applying targeted fixes, hold-out accuracy improved from 17.8% to 53.3%, and the most concerning negative comparison against a no_tool baseline disappeared. These results position Open-Rosalind as a biomedical-agent study with an explicit external-validity audit, rather than as a claim that protocol constraints alone guarantee superior performance.