In materials science, visual information is often the key evidence. If an agent is allowed to refer throughout the task to SEM/TEM microstructures, diffraction patterns, maps, spectral figures, and computed result figures, it may move closer to evidence-based interpretation than a simple text summary.
Experimentally, you can create a question set that includes images and measure whether the agent loses image references midway, answers based on the wrong figure, or can recover after failure. In particular, for internal review and analysis tools, you can also evaluate auditability by leaving visual evidence in place.