In materials automation, data parsing, model selection, simulation scheduling, experimental instrument control, and result analysis can easily drift apart, so it is useful to first create a benchmark that ties them together into a single executable objective. For example, one can connect a public dataset, a computational workflow, and a simple instrument emulator so that the agent completes the entire pipeline.
This approach is more suitable for bottleneck diagnosis than for immediate performance improvement. It is more realistic to organize where failures occur, which interfaces the model most often breaks, and what minimal tool chain can actually withstand the closed loop.