In materials research, please test a setup that assigns frequently separated tasks, for example literature summarization, experiment planning drafts, code writing, and data interpretation, to different models. If a human already knows the strengths of each model, it is better to start the selector with rules plus search rather than full automation.
At the beginning, keep only 2~3 models and record which model performs better for each question type. If you improve the selection logic based on those records, it will be easier to judge cost versus performance later when adding more models.