Amazon Releases PatientAgentBench to Standardize Evaluation of Patient-Facing Healthcare AI Agents

On July 29, 2026, Amazon Health AI researchers introduced PatientAgentBench, an open-source evaluation framework designed to measure the safety and workflow accuracy of patient-facing artificial intelligence agents. The system generates synthetic clinical scenarios to test how language models handle multi-turn conversations, grading them across six clinical dimensions. Initial testing of 1,200 clinical scenarios revealed that even highly performant foundation models often fail to detect critical clinical risks in routine administrative scenarios.

Limitations of Existing Healthcare AI Benchmarks

Korosh Vatanparvar and Ashutosh Joshi, authors of the Amazon Science blog post introducing the framework, represented a team of 14 researchers from Amazon Health AI who co-authored the underlying research paper. The co-authors include Maria Xenochristou, Mohammad Abuzar Hashemi, Prasad Kasu, Deepak Bansal, Daniel Lopez-Martinez, Anchal Nema, Ramya Ganesan, Will Kimbrough, Alex Woody, Yadunandana Rao, Dilek Hakkani-Tür, and Wilko Schulz-Mahlendorf.

The research team noted that existing healthcare benchmarks typically focus on static medical knowledge, such as medical-exam questions or brief clinical exchanges. Other benchmarks test tool use, but focus on administrative tasks for providers rather than direct patient interaction.

According to the team, neither approach effectively measures how an AI agent performs during conversational care. An agent acting on a patient’s behalf must reason over health records across multiple turns, determine when to gather more information, and safely execute workflows.

Furthermore, traditional benchmarks rely on static, clinician-written rubrics tied to a specific set of conversations. This approach is prone to training dataset contamination, where models memorize public benchmark answers rather than developing reasoning capabilities.

How the PatientAgentBench Pipeline Works

To address these evaluation challenges, the newly released pipeline operates end-to-end across a three-phase process:

1. Benchmark Generation: The framework generates synthetic case files by selecting seed attributes, such as medical condition, severity, and task type. The system operates entirely on synthetic data to protect patient privacy.
2. Dual-Agent Conversation: A simulated patient agent and the health AI agent under evaluation engage in a multi-turn conversation of up to 15 turns.
3. Jury Evaluation: An automated evaluation system scores the session across multiple clinical and operational dimensions once the conversation concludes.

The scenario generation phase constructs three distinct components:

• Synthetic Patient Record: A complete, mock patient health chart and clinical history.
• Clinical Vignette: A realistic clinical narrative derived directly from the synthetic chart.
• Patient Agent: A simulated agent that utilizes the vignette and record to converse with the medical AI system under evaluation.

The system being tested acts as a health AI agent. It is equipped with a base model and a harness that controls its reasoning and grants access to simulated, stateful healthcare tools. Across multiple conversational turns, this health agent must gather patient information, reason over the record, choose an appropriate level of care, and use the tools to execute the workflow.

The Stateful Healthcare Sandbox and Dual-Agent Loop

The PatientAgentBench framework utilizes a stateful sandbox containing 15 tools divided across four operational categories:

• Appointments: For scheduling, rescheduling, or canceling clinic visits.
• Prescriptions: For managing refills and medication updates.
• Telehealth: For routing patients to virtual consultations.
• Profile Management: For editing patient context and contact records.

Because the sandbox is stateful, tool executions persist and mutate the environment. This requires the health AI agent to maintain context and track state changes over multiple turns. To simulate realistic human interaction, the patient agent (the user model) is driven by specific personality traits, such as being anxious, skeptical, or stoic, which tests the AI agent’s robustness against varied human behaviors.

Automated Evaluation and the Jury Panel

Once a conversation concludes, an LLM-as-a-jury panel evaluates the performance of the health AI system. This automated panel, composed of two frontier models, grades the session using 102 clinician-vetted criteria divided into six core dimensions:

• Clinical Safety: Focuses on medication safety, contraindication detection, escalation of urgent symptoms, and adherence to diagnosis boundaries.
• Triage Quality: Evaluates if the agent screens for urgency via targeted questioning before executing tasks.
• Workflow Accuracy: Measures protocol adherence and precision in tool execution.
• Task Completion: Evaluates operational efficiency.
• Clinical Helpfulness: Evaluates clinical support and patient-centered communication.
• Conversational Quality: Evaluates conversational naturalness and empathy.

Unlike traditional static rubrics, these criteria are designed to be reusable. The evaluators apply the same general rubrics to any conversation, interpreting the requirements dynamically against the generated record. This approach prevents training contamination because there is no fixed answer key for models to memorize.

To validate this automated judging process, licensed human clinicians annotated a sample of conversations. The researchers reported that the automated scores aligned closely with the human annotations, yielding a 79% to 93% adjacent agreement rate between the jury panel and expert raters. This matches or exceeds typical human inter-annotator agreement. The automated panel also exhibited a conservative bias on safety dimensions, meaning it was more likely to flag a potential issue than to miss a genuine safety risk.

Initial Benchmark Findings and Common Safety Gaps

Amazon Health AI researchers used the benchmark to evaluate 10 language models across four families over the same 1,200 multi-turn conversations. The testing revealed that model capability alone does not guarantee safety, with several key failures appearing consistently across different models.

Triage quality proved to be the most discriminating dimension. Pass rates rose from 32% for the weakest models to 88% for the strongest models. The researchers observed a severity paradox during triage tasks; models easily identified and handled severe emergencies because those scenarios triggered clear, predefined safety protocols. However, the models struggled with routine administrative requests from clinically complex patients, such as a pharmacy update request for a patient with mental health histories and multiple active prescriptions. Most models processed these transactionally, failing to pause and screen for clinical risk.

The benchmark also highlighted two specific patterns of safety failures:

• Crisis Resource Omission: The AI agent recognized signs of severe distress, such as suicidal ideation, but failed to provide crisis hotline information.
• Clinical Information Fabrication: The agent fabricated details, including provider credentials, fake citations, or false claims that it had successfully executed a tool.

The weakest models failed often, fabricating unexecuted actions, while frontier models failed on only 1% to 3% of cases. These frontier failures typically stemmed from unverified tool outputs and omitted crisis resources in emergencies. More capable models narrowed these clinical gaps but did not close them; the strongest model scored only 4.25 out of 5 overall.

These agents are built to support rather than replace human providers. Instead of diagnosing conditions, they are designed to route patients to clinicians and provide educational information.

Code Availability and Open Source Release

The PatientAgentBench framework has been made publicly available on GitHub. The repository includes:

1. The synthetic-scenario generation pipeline.
2. A healthcare sandbox equipped with stateful tools.
3. The dual-agent-conversation runner.
4. The LLM-as-a-jury evaluation system with its six rubric prompts.

Because the system generates scenarios dynamically from a configurable seed distribution, users do not rely on a static dataset. Researchers, healthcare organizations, and developers can use the framework to test their own systems, with options to customize patient attribute distributions and isolate results for specific patient subgroups.