Optimizing genomics-aware clinical agents in precision oncology

Agentic systems are proposed for clinical decision support, yet unrestricted tool access can undermine accountability and safety in precision oncology, where recommendations must be grounded in evolving guidelines, regulatory labels, and patient-specific molecular evidence. We compare four architectures: (i) a standalone LLM, (ii) an unconstrained function-calling agent with over 200 biomedical tools, (iii) a ReAct agent over the same broad tool surface, and (iv) gSage, a workflow-orchestrated agent restricted to 20 curated domain tools backed by a versioned knowledge stack. We evaluate 375 clinician-authored questions across 41 cases spanning breast, non-small cell lung, prostate, ovarian, and colorectal/GI tract cancers using multiple frontier LLM backbones. A two-pillar evaluation, combining LLM-based rubric scoring with expert adjudication and an independent embedding-based content concordance analysis, yields convergent results. The workflow-orchestrated agent outperforms all alternatives on five rubric axes (guideline concordance, correctness, patient centricity, reasoning transparency, and safety) and achieves the highest embedding recall under both a biomedical (MedCPT, R = 0.877) and general-purpose (Snowflake Arctic, R = 0.803) model, significantly exceeding all full-sample arms (p < 0.05, Bonferroni-corrected). These findings show that principled workflow orchestration with constrained, curated tools, rather than maximal tool breadth, is what drives reliability and safety in genomics-aware oncology decision support.