A Multi-Agent Clinical Trials Framework to Automate Trial Center Operations

A team at Weill Cornell Medicine has outlined a multi-agent clinical trials framework in which a central AI coordinator delegates routine trial work to specialized agents while human experts keep the final say on clinical and regulatory decisions. The system, called ASTRA, is described in a Perspective published in npj Digital Medicine in partnership with Seoul National University Bundang Hospital.

ASTRA, short for Agent System for TRial Automation, is a design proposal rather than a report of results. The authors, Bjorn Redfors, Weishen Pan, Chen Lyu, Mario Gaudino, Fei Wang, and Hua Zhong, say formal validation will come later, through retrospective reconstruction of completed trials and prospective comparison. The paper sets out the architecture, the governance model, and the main barriers to adoption.

How multi-agent clinical trials are structured

Clinical Trial Centers, or CTCs, run the day-to-day work of multi-site studies. They handle protocol development, data management, quality assurance, safety oversight, outcome adjudication, statistical analysis, site monitoring, and laboratory support, usually split between Clinical and Data Coordinating Centers.

Ad

ASTRA uses a hub-and-spoke design. A central coordinator agent, which the authors call a virtual chief operating officer, issues work orders to team lead agents and tracks dependencies, status, and exceptions. It delegates all domain-specific execution and does not touch outputs such as DSMB narratives or randomization schedules. Below it sit ten teams covering protocol development, data management, quality assurance, biostatistics, a Clinical Events Committee, a Data and Safety Monitoring Board, site operations, site monitoring, patient interactions, and core laboratories.

Teams communicate through a standardized messaging protocol rather than free text. Each message is a structured object with a task identifier, sender and recipient, data pointers, a status field, an escalation flag, and a human-approval status. Agents pull from a shared, version-controlled knowledge base that holds protocols, statistical analysis plans, case report form schemas, and adjudication rules, and they connect to outside systems through the open Model Context Protocol.

Three task classes

ASTRA is modular rather than tied to one level of automation. The authors split work into three classes. Human-led tasks need clinical judgment, ethical discretion, or formal sign-off, such as a chief statistician approving an interim analysis or a quality lead signing a corrective action. System-assisted tasks prepare or structure material for review, like assembling a safety packet for a monitoring board. Bounded agent-executed tasks are routine and rule-governed, such as validating case report forms against a schema or flagging out-of-range data fields.

The cost case

The motivation is cost. NIH-funded studies often run between $5 and 20 million, and US industry Phase III trials typically cost $12 to 50 million or more. CTC operations absorb 25 to 45 percent of that spend, sometimes more than half for complex multinational trials, with data management and quality oversight alone accounting for 15 to 25 percent. Coordinators spend more than 260 hours over six months on query resolution and source data verification, and staff turnover costs exceed $250 million a year.

Regulatory guardrails and hallucination controls

The framework arrives as regulators move toward AI in trial operations. The FDA has launched Elsa, a large language model tool for protocol evaluation, adverse-event summarization, and inspection targeting, and its January 2025 draft guidance on AI in regulatory decision-making sets out a risk-based approach to AI credibility. ASTRA leans into that: a time-stamped, immutable audit trail built for 21 CFR Part 11 and GDPR, strict separation of blinded and unblinded work, and role-based access.

The authors treat hallucination as a first-order hazard to contain by design. Agents operate in tightly scoped roles, grounded in retrieved source material and constrained by deterministic schemas and rule-based validators. Outputs that fail validation are escalated rather than corrected automatically, and models, prompts, and tools are version-locked for the life of a trial unless formally revalidated. Routine drift audits may become part of trial oversight.

A staged path to validation

Because ASTRA is not yet validated, the authors propose three steps. First, retrospective simulations would reconstruct completed trials and compare agent outputs, timelines, and escalation events against the historical record. Second, shadow-mode deployment would run agents in parallel with normal operations without influencing the trial. Third, low-risk domains would be evaluated live using measures such as task completion time, query-resolution time, protocol-deviation detection, data completeness, cost per task, audit-trail completeness, and user workload.

The goal is a hybrid model in which bounded agents handle routine operational work while human experts keep judgment, oversight, and accountability. That is a long way from the paper case report forms, double data entry, and manual validation that still dominate many data coordinating centers.

Source: npj Digital Medicine, “Towards a multi-agent Clinical Trial Center framework to support modern clinical trials” (Redfors et al., 2026).