Adaptive Influence Graphs debug failed multi-agent runs the way observability tools work, setting a new state of the art on Who&When

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 25 August (arXiv 2608.24361), this work starts from the observation that engineers do not debug distributed systems by reading raw logs end to end, they use observability tooling organized around components, actions, and dependencies. AIGs is a two-stage agentic framework that first converts a failed multi-agent trace into a structured graph, then navigates that graph to localize the critical error. Across multiple models, richer trace representations consistently improve failure attribution, with adaptive graph construction plus agent-directed traversal strongest, and the method sets a new state of the art on Who&When, the standard multi-agent failure attribution benchmark. The finding that attribution depends on trace representation and not just diagnosing model quality is directly actionable for anyone instrumenting an agent pipeline.