Jichi Medical University built LLENS, an AI agent platform that runs on hospital-owned servers, sending no patient data outside the building. The system answers clinical questions by searching internal documents, ordering steps, and returning sourced answers to doctors and nurses. It ran on Llama 3.1 70B, a large language model the hospital hosted itself. The team tested two ways the agent could find information. In one mode it searched documents once before answering. In the other it searched again and again, refining its query as it went. The second approach scored higher on accuracy across all clinical categories the team measured. That finding matters for designers: how an agent gathers information mid-task changes what the human receives at the end. The study also shows what trust looks like in a high-stakes setting. Clinicians could see which source each answer came from. The platform logged each step. No answer went to a patient without a doctor reading it first. Visible sources, full logs, and a required doctor review form a working model for agent design in any domain where a wrong answer causes real harm.