benchmark.darvinyi.com
← Back to Agent Evaluations
Microsoft Research2026

HealthAgentBench

Microsoft Research suite of 54 agentic healthcare tasks across 7 real clinical/biomedical workflows (EHR auditing, X-ray report correction, clinical trial matching, pathology, CT), scored by task-specific verifiers with cost/time tracking.

Tasks54
Primary MetricTask Success Rate
Avg Human TimeRanges by task: 15–60 min (X-ray correction) to >8 hours (clinical trial matching)

Overview

HealthAgentBench is a unified benchmark suite from Microsoft Research (lead authors Qianchu Liu, Sheng Zhang, Guanghui Qin, Hoifung Poon and collaborators) that evaluates frontier AI agents on realistic, end-to-end healthcare workflows rather than static medical QA. Each of its 54 tasks, spanning 7 categories (EHR format conversion, X-ray report correction, CT abnormality classification, pathology tumor-area selection, clinical trial matching, EHR data-quality auditing, and EHR event modeling), drops an agent into a containerized terminal environment built on the Harbor framework, where it must inspect real clinical data (drawn from credentialed sources like MIMIC-IV, MIMIC-CXR, and TREC Clinical Trials), use tools, and produce a verifiable output judged by a task-specific automated verifier or LLM-as-judge.

The benchmark was built to address the gap between narrow medical QA leaderboards and the messy, multimodal, long-horizon nature of actual clinical and biomedical work — covering modalities from 2D X-rays and 3D CT volumes to gigapixel pathology slides and structured EHR data, with documented expert time-to-solve reference points (e.g. clinical trial matching takes experts over 8 hours; X-ray report correction 15–60 minutes).

Results show current agents remain far from reliable: the best agent, Codex GPT-5.5, solves only 42% of tasks overall (at roughly $0.70/task), with EHR/text-based tasks averaging 49% success but medical imaging tasks averaging just 17%. The paper also surfaces a notable cost/performance disconnect — Claude Code agents were both the slowest and most expensive (up to $4.80/task) yet scored below cheaper competitors, largely due to producing ~2x more output tokens per trial.

How It Works

Task setup, inputs, outputs, and evaluation.

SetupAgent is placed in a Docker-based Harbor container with credentialed access to real clinical data sources, isolated from verifier gold labels.
InputA concrete clinical or biomedical problem instruction (e.g. audit EHR records for data-quality issues, correct an X-ray report, match a patient to a clinical trial).
OutputA structured artifact (corrected report, converted data format, classification, matched trial ID) written to the container filesystem.
EvaluationTask-specific hidden verifiers compare output against gold labels; imaging/text tasks use LLM-as-judge with majority vote.
MetricBinary/graded Task Success Rate, with per-task dollar-cost and time tracked as secondary axes

Results

Model performance on this evaluation.

#ModelScore
1
Gpt 5 5
42%
2
Claude Opus 4.7
36%

Key Findings

  • Best agent (Codex GPT-5.5) solves only 42% of tasks overall; EHR/text tasks average 49% success vs. just 17% for medical-imaging tasks.

  • Claude Code agents were both the slowest and most expensive (up to $4.80/task) yet underperformed cheaper models, producing ~2x more output tokens per trial.

  • EHR data-quality auditing success jumps from 42% to 97% when search clues are provided, showing large search space (not reasoning) is the primary bottleneck.

  • 9 of 10 agents achieve 100% success on EHR format conversion, showing strong performance on pure text/structured-data transformation but sharp drop-off on multimodal imaging.

What Makes It Unique

  • Spans the full patient-journey and data-modality range (2D/3D imaging, whole-slide pathology, structured EHR, free text) in one suite, unlike single-modality medical benchmarks.

  • Reports per-task dollar cost and wall-clock time alongside success rate, enabling Pareto-frontier cost/performance analysis.

  • Uses real credentialed clinical datasets (MIMIC-IV, MIMIC-CXR, TREC Clinical Trials) with anti-cheat measures (scrubbed IDs, credential isolation, no web access) to prevent contamination.

Controversies & Caveats

Requires credentialed access to PhysioNet datasets (MIMIC-IV/MIMIC-CXR), limiting fully open reproducibility.

Some task categories (CT classification, pathology) have small task counts, so per-category scores carry high variance.

Links