AI agents modeled on human reasoning may accelerate scientific discovery, authors argue

The article argues that AlphaFold’s success depended on a rare combination of standardized data and replicable experimental methods, conditions unlikely to be met in most scientific fields. It notes the Protein Data Bank required 53 years and an estimated $21 billion to assemble, and that replicable measurement standards are uncommon outside fields like genomics and weather forecasting.

The authors contend that for most open scientific questions, AlphaFold-style breakthroughs are not imminent, and instead propose AI agents as a more practical near-term path. These agents integrate large language models with tool access to mimic the iterative, hypothesis-driven process of research, including literature review, critique, experimentation, and refinement.

The piece highlights Google DeepMind’s AI Co-Scientist, which autonomously generated and tested hypotheses about antibiotic resistance, reaching a correct conclusion that had previously taken human researchers a decade to establish. The system used multiple sub-agents to draft, critique, rank, and refine hypotheses without prior exposure to the relevant peer-reviewed paper.

The authors acknowledge current limitations of AI agents, including hallucinations, inconsistent judgment, and constraints on memory and runtime, but argue these technical barriers will diminish over time. They also suggest agents could structurally improve reproducibility by automatically logging every step of their process, creating an auditable trail for replication.

The article frames agents as a foundational shift toward modeling the human process of discovery rather than applying narrow, data-intensive solutions to limited problems.

It concludes that while data-driven models like AlphaFold will remain impactful in specific domains, AI agents offer a more generalizable and scalable approach to accelerating science across diverse fields.