Preprint proposes hierarchical server architecture for agentic science

The preprint describes a hierarchical architecture in which “secretary” agents negotiate and select services across cloud, edge and high-performance-computing providers. The authors report probing 51 real or simulated providers and achieving 87.71% negotiation accuracy across 19,973 negotiation and 6,952 selection simulations. The work is intended to support distributed scientific workloads, including the Genesis Mission.

Impact: the paper offers a concrete coordinator pattern for heterogeneous compute discovery, policy and delegation. Its results are author-reported simulations; there is no peer review, independent reproduction or demonstrated production reliability.