Telling an Agent the Knobs Are Architecture Beat Anonymous Variables by 12.3%, but a Critic Loop Erased the Gap

arXiv 2609.19387·medium signalXBlueskyLinkedInCopy link

arXiv 2609.19387 (submitted 16 Sep 2026) separates whether an agent reasons about hardware from whether it merely searches well, by handing the same agent the same 15-dimensional accelerator space twice: once as named architectural knobs with simulator counters, once as anonymous variables on [0,1], with evaluator, legal space and reachable optima held identical. On a nine-kernel FP16 GEMM basket the informed architect beat a modeled H200 by 5.4% and its blind counterpart by 12.3% on average, using 70.1% fewer simulator calls. A critic loop recovered most of that gap for the blind agent and bought the architect nothing, so domain knowledge and structured critique act as substitutes; the authors flag five to six runs per condition on one modeled accelerator as preliminary.