Developer builds multi-agent AI 'courtroom' that debates contested questions instead of giving one answer

A developer has built RabbitHole, a multi-agent AI system using LangGraph that pits different AI personas — such as a state advocate, privacy activist, and compliance officer — against each other in a structured debate before a judiciary node delivers a verdict with a confidence score. The project was motivated by frustration with standard RAG systems, which return a single confident answer even for genuinely complex or contested questions like legal and policy issues. RabbitHole uses a nested two-graph architecture: an outer Courtroom graph managing the debate flow and an inner RAG sub-graph handling hybrid search, reranking, and hallucination checks. A key engineering lesson from the build was that structural constraints — such as limiting the number of debate perspectives — should be enforced in the graph's state schema rather than relying on prompt instructions to the model. The project is not yet deployed due to cost constraints, but the developer describes it as functionally complete end to end.