We present a closed‑loop autonomous workflow built on a coding‑agent stack, executing synchrotron X‑ray measurements on nanoparticle superlattices. The system follows evolving human‑defined goals and logs complete human‑agent and agent‑agent dialogues for post‑hoc audit. The audit reconstructs decision provenance and the split of cognitive labor: 61%‑67% of all messages contain interpretation or judgment, while methodological and scientific input rises from 46% in the first run to 56%‑61% in later runs. Error corrections involve multiple participants in 59%‑68% of cases, and roughly half of the decisions incorporate another participant’s input. These findings offer concrete design guidance for multi‑agent scientific platforms and inform model training and accessible autonomous experimentation.
Review: By providing a granular audit of human‑agent interaction, the study quantifies cognitive allocation and sets a reproducible benchmark for future large‑scale multi‑agent experiments.