In this work sixteen independently initialized sessions employed the same model‑harness setup, accessing a frozen library of 12,499 metal‑organic frameworks (MOFs) with a methane‑storage objective, a fixed protocol and a one‑week computational budget. Despite identical inputs, the agents diverged into four screening strategies covering roughly 100 to 5,000 structures, while eight sessions also generated 2,253 hypothetical candidates. All agents converged on the same materials frontier near 200 cm³/cm³ of pore volume. An independent analysis of the database’s porous region confirmed that the nine best structures were all reported by the agents. Enforcing additional checks on half of the agents raised fresh‑run reproducibility from one‑of‑eight to eight‑of‑eight, yet it did not improve the validity of the conclusions because fifteen of the sixteen agents still selected the same audit‑excluded entry—an incomplete structure missing anions that artificially inflated pore volume. Thus replicated agents reveal both robust findings and common‑mode errors arising from shared inputs.
Review: The study demonstrates that autonomous material discovery can follow diverse pathways yet arrive at a consistent frontier, highlighting both the strength of the underlying algorithm and the risk of systematic bias from shared data artifacts.