Medical AI is moving from answering questions to doing work: processing prior authorizations, working a specialist consultation across dozens of tool calls, auditing a chart against a clinical standard. A single-turn benchmark cannot show if an agent completed the workflow it was given, effectively planning and using tools. Current results show the gap is large: on HealthAdminBench, agents complete 83% of individual subtasks but only 36% of workflows end to end. On PhysicianBench, the best of 13 agents succeeds on 46% of tasks. on CliniCARE-Bench, every one of 16 systems tested committed to an answer even when the patient record doesn’t support one.”This session walks through the three agentic benchmarks now integrated into the open-source MedHELM framework. They all score what the agent actually did rather than what it said it would do:
- HealthAdminBench (Stanford, Shah Lab): 135 multi-step workflows across prior authorization, appeals and denials, and DME ordering, decomposed into 1,698 verifiable subtasks and executed by computer use agents in a browser.
- PhysicianBench (Stanford, HealthRex lab): 100 long-horizon tasks adapted from real primary-care-to-subspecialty consultations across 21 specialties, averaging 27 tool calls per task.
- CliniCARE-Bench (Scale AI): 25 clinician-authored audit scenarios instantiated as 750 patient cases over MIMIC-IV data, with two indeterminate verdict classes that make principled abstention part of the score.
- How the extended MedHELM harness evaluates agents that plan, call tools, and work across many turns.
- How to run the benchmarks yourself: install, configure your model, and execute a workflow, and populate a leaderboard.
- Why one open-source platform for accuracy, agentic reliability, safety, and bias testing beats a patchwork of tools, and how the same suites run as a CI/CD gate with Gatekeeper and as a production monitor with Guardian.
