LLM-as-an-Investigator: Evidence-First Agentic Reasoning for Robust Interactive Problem Diagnosis
The LLM-as-an-Investigator methodology reframes LLMs from direct answer generators to interactive hypothesis investigators for technical troubleshooting. It uses a Solution Investigator Agent that manages competing hypotheses, asks targeted clarification questions, and updates probability scores based on evidence. This approach reduces premature commitment to wrong solutions by forcing the model to distinguish between causes through structured investigation.