Introspecter: Diagnosing Multi-Agent LLM Debate Failures

Multi-agent debate systems often fail through error propagation, social conformity, and premature consensus, yet existing evaluation pipelines provide limited support for analyzing how such failures emerge across interaction trajectories. We present Introspecter, a framework for intervention-based analysis of failures in multi-agent reasoning systems under both collaborative deliberation and argumentative debate protocols. Introspecter provides: (1) a configurable debate environment for LLM and human participants with trace-level annotation of suspected failure points, (2) a counterfactual replay mechanism that allows intervention on intermediate agent states and resimulation of downstream interactions to test causal hypotheses about failure propagation, (3) an argumentative debate mode with stance-constrained and citation-grounded argumentation for studying adversarial failure modes, and (4) analysis and visualization tools for exploring debate trajectories and diagnosing failures. We release Introspecter as an open-source platform for reproducible analysis of multi-agent reasoning failures and protocol-dependent failure dynamics.