Agent evaluations run against the real agent patients would encounter

LangChain says every evaluation runs through the same graph against the real agent, so tests score what a patient would encounter. Tests range from single-reply fixtures to simulated multi-turn conversations with LLM patient personas. LangChain says LangSmith records conversation traces for reviewer annotation, holds datasets and experiments for comparisons, and receives scores from triage and improvement workflows.