Teaching Model Context Protocol, Retrieval-Augmented Generation, and AI Agents to a Multidisciplinary Hospital...

Hospitals worldwide need to upskill their workforce in advanced AI technologies; yet, published guidance on how to design and deliver such training, particularly in agent-level tools like retrieval-augmented generation (RAG) and the model context protocol (MCP), remains virtually absent.

To describe the design, implementation, and lessons learned from an 8-week, 56-hour intensive generative AI training program for a multidisciplinary hospital workforce, drawing on both quantitative outcome data and participants' own reflections on their learning experience.

The program was delivered on-site at Asan Medical Center with simultaneous online broadcast to 2 regional affiliate hospitals. The curriculum was built around the premise that MCP and AI agents would become the foundation of health care AI use, allocating 37% (11.5/31 hours) of on-site instructional time to MCP, and 71% (22/31 hours) to hands-on practice. Participants progressed from foundational concepts through RAG and MCP to team-based capstone projects, supported by funded AI tool subscriptions, a dedicated internal cloud platform, and 3-6 hours of weekly mentoring per team. A pre-post survey (pre: n=83; post: n=64) evaluated outcomes across Kirkpatrick levels 1-3, complemented by thematic analysis of open-ended reflections on self-perceived growth.

The technologies that received the greatest curricular investment were associated with the largest self-efficacy differences (MCP: d=1.57; overall effect: r=.574), and participants most frequently cited MCP and RAG when describing how abstract concepts "became concrete and actionable." Non-IT professionals, clinicians, health information managers, researchers, and administrative staff showed consistently larger gains than IT specialists; several reported coding for the first time through vibe coding, challenging the assumption that advanced AI training requires technical backgrounds. Despite significant overall gains, a knowledge-practice gap persisted: job-specific competency remained below the scale midpoint, though participants spontaneously reported generating workplace application ideas. Curriculum pacing was rated lowest despite high overall satisfaction (4.03/5), signaling that even 56 hours may progress too quickly for mixed-expertise cohorts. Capstone projects with dedicated mentoring received the highest satisfaction ratings; 11 of 12 teams presented functional prototypes, and one has since entered active pilot use in clinical departments ahead of planned hospital-wide deployment.

To our knowledge, this is the first program to teach 4 agent-level generative AI technologies, MCP, RAG, LangGraph orchestration, and AI agent design, to both IT and non-IT hospital staff. This program suggests that transforming a multidisciplinary hospital workforce into AI-capable professionals is achievable through intensive, hands-on training centered on agent-level technologies, and that capstone projects with dedicated mentoring can serve as a pathway from classroom learning toward institutional AI adoption. The knowledge-practice gap highlights the need for posttraining support structures to translate self-efficacy gains into sustained workplace practice.