Every few decades, predictions emerge that science has reached its end. In 2024, Demis Hassabis and John Jumper of Google Deep Mind were awarded part of the Nobel in chemistry for Alpha Fold, which predicts protein structures by learning from thousands of experimental shapes. While this success sparked a wave of investment, Alpha Fold's success depends on massive, specialized datasets like the Protein Data Bank—a resource requiring 53 years of international cooperation and $21 billion to assemble.

Most scientific fields lack such consistent, scalable data due to experimental variability. Consequently, the true acceleration of science may come from AI agents rather than pure data-driven models. Unlike Alpha Fold, which applies a powerful approach to a limited question, agents act as generalists that digitally model the human process of discovery by reasoning through uncertainty.

By using large language models to access digital and physical tools, agents can mimic the iterative research process. Tools like Google’s AI Co-Scientist can draft, critique, and refine hypotheses, often reaching correct conclusions faster than traditional wet-lab work. As these agents mature, they promise to solve the reproducibility crisis and amplify scientific memory through automated, standardized logging of every research step.