If the tool list in a materials laboratory has grown large, you can try an experiment that groups simulation codes, structure parsers, database query tools, plot generation scripts, and optimization routines into one skill library and then first extracts only the top candidates. This can reduce the problem of the agent becoming confused by too many choices.
At first, it is safer to start as a tool-selection recommender rather than full automation. For example, you can compress “For this question, which is appropriate: DFT calculation, XRD peak fitting, or literature search” down to only the top 3, and then validate a workflow in which a human makes the final choice.