In materials automation, calculation submission, database lookup, postprocessing, visualization, and experimental equipment control are intertwined, so an agent may become unstable if it selects tools one step at a time. Therefore, one possible approach is to have it select bundles such as a DFT submission tool, a parsing script, a property DB lookup, and a plot generator.
Initially, it is realistic to start with a limited internal stack of about 20~50 tools and train the system to learn which combinations are often needed for which task types. Later, evaluation could be based less on “which single tool to use” and more on “which tool bundle is needed to finish the next 10 minutes of work.”