In materials research labs, DFT, MD, synthesis, characterization, and post-processing scripts are often scattered separately, so it may be worth first trying a way to wrap the input and output of each stage as typed tool calls. If the agent writes and reads composition, structure, conditions, and measured values in a normalized table instead of directly reading files, it becomes easier to track experimental history.
Next, it is better to choose and automate only one narrow loop. For example, one could build a loop such as candidate composition generation→DFT calculation submission→result collection→next candidate recommendation, and store all intermediate outputs in a queryable schema. This would make it possible to separate the agent’s reasoning errors from data omissions.