Agentic chemistry needs evaluation, not only demos

Agentic systems in chemistry are exciting because they can bridge interfaces, tools, and scientific intent. But demos alone are not enough.

Useful scientific assistants need to be evaluated on whether they finish realistic tasks, whether they choose sensible intermediate steps, and whether their outputs remain scientifically defensible when workflows become more complex.

That is why benchmark design is central to this area, not an afterthought.