Samsung Medical Center · KOHI Medical-AI Workforce Training, Advanced 3-Hour Session
The advanced session of a government-funded workforce training program. Three hours of theory on agents and Tool Use, in a room where three-quarters of the audience were not physicians. The question that stayed longest was whether I use remote control.
Overview
- Date: Wednesday, 15 July 2026, 18:30–21:30 (3 hours)
- Venue: Samsung Medical Center, Ilwon Campus
- Host: Ministry of Health and Welfare · KOHI, "2026 Medical-AI Workforce Training for Health Professionals," Samsung Medical Center track
- Course: Medical-AI agents and Tool Use, theory and deployment strategy (intermediate, level 3 advanced)
- Attendees: ~60. Administration 32%, nursing 27%, allied health 23%, a small share of physicians (mostly faculty)
- Format: Theory-led, three hours. No participant hands-on; demos ran as three pre-recorded videos
Why this session
Most of my earlier talks put physicians in front of a laptop and had them run things. This one was the opposite. Three-quarters of the room were not physicians, nobody executed anything themselves, and all three hours were theory.
I stripped out the clinical vocabulary and rewrote the examples in language shared across job roles. "You don't need to know how to code" moved to the front. The demos moved from live execution to three pre-recorded videos, because I did not want a three-hour session in front of sixty people hanging on the network and the clock.
What it covered
The spine was a five-step ladder: prompt → context → skill → MCP → harness. A chatbot gives you an answer; an agent gets the work done. Tool Use bridges them: the model starts calling search, code, files, and databases directly.
Forty minutes went to case studies. After three hours of theory, what people are left with is "so does this actually work?" Showing results I ran and verified myself was how I answered that.
- MedSci Skills (arXiv:2606.09500) — gates that verify an AI-drafted manuscript before a human reads it.
- MeducAI (npj Digital Medicine 2026) — a governance-first pipeline where the LLM generates and code decides.
The last thirty minutes were safety and limits: context hygiene, de-identification, human review, and where a hospital should start.
The question that stayed
In the Q&A someone asked: "Do you use remote control?" Meaning, can you drive Claude on your home machine from your phone.
I said no. Many people expect that once you start vibe coding with AI, efficiency rises and the work shrinks. My experience is the opposite. As productivity goes up you take on more work, and your head gets busier. So I deliberately never set up remote access, because when I am not at my computer I want to be resting. The people who did set it up are working from the bathroom and from vacation.
Separating work from rest is going to be one of the important skills.
Looking ahead
The session was recorded and stays with the trainees. Two hands-on sessions follow in August.
- 19 August — Organizing research material and knowledge management with AI
- 26 August — The manuscript pipeline
Twenty people, lab PCs, each bringing their own data. They will run for themselves what they only watched on screen this evening. Thanks to the organizers for making the space, and to everyone who stayed all three hours. Our experiment continues.