From Supergraphs to Small Teams: Contrastive Attribution for Multi-Agent Topology Search

LLM-based multi-agent systems depend on their communication topology, yet task scores alone hide which agents help or hurt. We introduce contrastive-attribution topology search, an offline meta-agent that samples subgraphs of a supergraph, assigns signed per-agent credit from execution transcripts, and prunes weak edges into a small fixed team. On BigCodeBench, the selected team scores 5.3 points above budget-matched score-only REINFORCE while using 54% fewer tokens. In a poisoned-pool stress test on HotpotQA, it excludes all poisoned agents across three seeds, whereas random search recruits $2.0 \pm 1.0$ of them at comparable score. This supports per-agent credit as an optimization and audit signal.