In realistic clinical settings where agents have access to tools and networks, their diagnostic accuracy could rise from 20% to around 50–60%.
