PatientAgentBench: A Benchmark Framework for Evaluating Patient-Facing Health AI Agents

PatientAgentBench: эталонная платформа для оценки медицинских ИИ-агентов, взаимодействующих с пациентами
Maria Xenochristou, Alex Woody, Dilek Z. Hakkani-Tur, Korosh Vatanparvar, Ashutosh Joshi, Mohammad Abuzar Hashemi, Prasad Kasu, Deepak Bansal, Daniel Lopez-Martinez, Anchal Nema, Ramya Ganesan, Will Kimbrough, Yadunandana Rao, Wilko Schulz-Mahlendorf
2026-07-28

LLM-as-a-JuryPatientAgentBenchagentic healthcare evaluationclinical safetypatient-facing health AI agents
Health AI is evolving from answering questions to agentic systems that converse with patients, reason about health records, and act on their behalf. Primary care guards against diagnostic errors and unsafe care; agents assisting in this domain warrant evaluation against the same risks. Current benchmarks focus on medical knowledge, assessed through isolated question-answering or clinician-facing tasks. PatientAgentBench benchmarks patient-facing agentic healthcare; it evaluates a foundation model, wrapped in an agent with a sandbox of healthcare tools, conversing with a simulated patient. Each conversation is scored by an LLM-as-a-Jury across six dimensions via over a hundred conversation-agnostic, clinician-grounded criteria. To validate alignment, licensed clinicians annotated shared conversations, yielding 79-93% adjacent agreement between jury and expert raters, on par with or exceeding clinician inter-rater agreement. We benchmarked 10 models across four families on the same 1,200 scenarios and found clinical gaps. Triage quality is the most discriminating dimension: pass rates rise from 32% for the weakest models to 88% for the strongest, with agents often acting on administrative requests without clinical screening. Clinical safety and workflow accuracy follow the same pattern: the weakest models fail often, fabricating unexecuted actions, while frontier models fail on only 1-3% of cases, from unverified tool outputs and omitted crisis resources in an emergency. More capable models narrow these gaps but do not close them; the strongest scores only 4.25 of 5 overall. These failures surface only in sustained, tool-using conversations against realistic patient records, confirming that static benchmarks are insufficient as healthcare agentic systems gain autonomy. We release the framework as a reproducible, clinician-validated evaluation standard to help the field close this gap.
1
Across 10 models and 1,200 scenarios, triage was most discriminating, with pass rates ranging from 32% for the weakest models to 88% for the strongest.
2
More capable models reduced but did not eliminate safety gaps: the strongest achieved only 4.25 out of 5 overall, with failures involving unverified tool outputs and omitted emergency resources.
3
PatientAgentBench evaluates patient-facing healthcare agents through sustained conversations with simulated patients, realistic records, and sandboxed healthcare tools.
4
The benchmark scores agents across six dimensions using over 100 clinician-grounded criteria and an LLM jury, achieving 79–93% adjacent agreement with licensed clinicians.
5
The findings show that static knowledge benchmarks miss safety failures emerging during realistic, tool-using healthcare conversations; the authors release a reproducible, clinician-validated framework.
6
Weaker agents frequently acted on administrative requests without clinical screening and fabricated unexecuted actions, while frontier models still failed 1–3% of cases.

patient-facing agentic healthcare systems conversing with simulated patients and interacting with realistic patient health records through healthcare tools

clinical safety, triage quality, workflow accuracy, and related failure modes in sustained tool-using patient conversations

Publication Details
Publication Date
2026-07-28
Journal
Publisher
ISSN
Cited by
4
Access Type
Author Information
Authors
Maria Xenochristou
Alex Woody
Dilek Z. Hakkani-Tur
Korosh Vatanparvar
Ashutosh Joshi
Mohammad Abuzar Hashemi
Prasad Kasu
Deepak Bansal
Daniel Lopez-Martinez
Anchal Nema
Ramya Ganesan
Will Kimbrough
Yadunandana Rao
Wilko Schulz-Mahlendorf
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%