AI Research

MedAgentBench and the Clinical AI Frontier: Stanford’s Benchmark for Healthcare Agent Safety

Stanford University’s benchmark suite exposes a critical gap between medical knowledge and clinical execution: frontier AI models achieve 69.67% success in realistic EHR workflows, until systematic architectural redesigns — including extractive memory, tool abstraction, and mandatory human-in-the-loop gating — push reliability to 98%. From