AFANet: A Lightweight GNN Matches Fine-Tuned LLMs at Blaming the Right Agent in a Failed Multi-Agent Run

arXiv 2608.18575·medium signalXBlueskyLinkedInCopy link

Agent failure attribution, identifying which agent in a failed multi-agent trajectory broke and how, is currently done with prompting, fine-tuning or agentic pipelines that burn long-context tokens. AFANet (arXiv 2608.18575, 2026-08-19) models trajectories through step-level semantic signals and agent-level relationships in a graph instead, and with far fewer parameters and near-zero inference cost it matches or beats LLM-based baselines including in-domain fine-tuned models. It holds up across different GNN architectures and improves further with cheap test-time adaptation on the out-of-distribution benchmark, suggesting observability tooling for agent fleets does not need an LLM in the loop.