Agentic AI

Grand Rounds October 9, 2026: An End-to-End Agentic AI Platform to Automate Clinical Trials: A Proof-of-Concept and Evaluation Study (Christoph Hornik, MD, PhD, MPH; Moulik Shah)

Posted on by Sophia Ramirez

Speakers

Christoph Hornik, MD, PhD, MPH
Duke Clinical Research Institute

Moulik Shah
Founder and CEO
Maxis AI

Keywords

Agentic AI; Clinical Research; Clinical Trials; Chatbot; AI Training

Key Points

  • Project Loom is a proof-of-concept study aimed at evaluating agentic artificial intelligence (AI) workflows end-to-end. The research team utilized pre-built agents with specific roles: writer, reviewer, builder, recruiter, and reporter. The agents were assembled on one platform to execute the trial. Human oversight took place at each step.
  • The agents were successful in 4 out of 5 operational domains and completed the work in 182 working hours. The presenters outlined several changes that they are making in response to failures in eligibility screening, chatbot hand-offs, and others.
  • The presenters concluded that end-to-end agentic AI execution is feasible and fast. This technology is meant to empower researchers and expedite their work, not to replace them. Training and testing of AI agents and workflows will be critical.

Discussion Themes

Document drafting, case report form generation, and backend data extraction are considered ready for live use today under human review. Direct participant-facing interactions and patient safety monitoring are viewed as the highest-risk areas requiring further refinement and testing.

The platform operates within institutional firewalls using strict role-based access control, bounded scopes of execution, and an observability layer that tracks every computation and handoff.

Resolving the EHR identification and screening issues is technically feasible by profiling unstructured clinical text alongside structured data and extending formal handoff contracts to third-party tools.

The term "Project" signifies that the system is an evolving, iterative initiative rather than a static commercial product.

Grand Rounds August 28, 2026: Evaluating Clinical Agentic AI (Haider Warraich, MD)

Posted on by Sophia Ramirez

Speaker

Haider Warraich, MD
Program Manager
ARPA-H (Advanced Research Projects Agency for Health)

Keywords

AI; Outpatient Care; Agentic AI; Heart Failure; Clinical Care

Key Points

  • The current state of outpatient care is sporadic and reactive. Low-value care leads to excess costs and harms, and clinical resources are wasted on administrative tasks. Clinical agentic AI could provide patients with round-the-clock access to specialty care and unlock the full potential of the Electronic Health Record and wearables.
  • ADVOCATE is envisioned as a patient-facing agentic AI system that can provide 24/7 clinical care as an extension of the care team. It would include administrative and wellness functions, like care navigation and diet and exercise recommendations; advanced decision support in regard to treatment and diagnosis; and high-risk functions, like prescription management. The ARPA-H team hopes it will transform advanced cardiovascular disease management.
  • Their current focuses are the development of an agent for patients with heart failure; the development of an independent supervisory agent to ensure continued agent accuracy and patient safety, particularly after clinical deployment; and recruitment of health systems for the design, development, and deployment of the technologies.
  • The clinical evaluation strategy for ADVOCATE will involve 3 phases: Human Factors Evaluation and simulated virtual patient testing; shadow-mode evaluation in clinically deployed settings; and a stepped-wedge multicenter cardiovascular outcomes trial.

Discussion Themes

While underlying foundation AI models possess broad medical knowledge, their autonomous actions must be strictly constrained to their labeled, intended use to maintain safety.

Heart failure was selected – as opposed to upstream prevention – in part because it is characterized by a highly algorithmic care path, a high event rate that generates clinical study data quickly, and strong financial incentives.

While the program initially targeted non-inferiority compared to a cardiologist, the metric was updated to require a superiority analysis against usual care.