A multi-agent LLM system runs fault management across millions of optical links in production AI datacenters, cutting fault incidents over 60%

arXiv·medium signalXBlueskyLinkedInCopy link

Reported 24 August (arXiv 2608.23145), this is described as the first LLM-powered multi-agent system deployed for autonomous fault management across millions of optical links in production AI datacenters, not a simulation. Refined via supervised fine-tuning plus continuous memory evolution, it reports 97.7% F1 and over 60% reduction in fault incidents on a ten-week field data evaluation, outperforming the state-of-the-art LLM baselines it was compared against. The paper is a short demonstration report rather than a full system description, so the architecture detail is thin, but the ten-week production window at that scale is the rare thing here.