Artificial intelligence is already changing how scientists work, but the next leap may not come from bigger datasets alone. The stronger near-term path could be AI agents: systems that use reasoning and tools to imitate the back-and-forth process of research.
The argument is not that data-driven breakthroughs like AlphaFold are unimportant. It is that AlphaFold depended on conditions most areas of science do not have, and may not have for a long time.
AlphaFold proved the promise and the limit
In 2024, Demis Hassabis and John Jumper of Google DeepMind received part of the Nobel in chemistry for AlphaFold, a neural network that predicts the three-dimensional structures of proteins. The achievement mattered because the protein-structure problem had resisted broad solutions for half a century.
AlphaFold became a symbol of what AI for science might become. Hassabis and his team called it “the template for how AI can accelerate all of science to digital speed.” After that success, startups working on foundation models for biology, chemistry, and materials discovery attracted billions of dollars.
But AlphaFold was possible because one field had something unusual: a large, carefully assembled base of experimentally validated protein structures. The Protein Data Bank contained roughly 170,000 such structures, built through 53 years of international scientific cooperation and, by a recent estimate, roughly $21 billion worth of experimental work.
That kind of foundation is hard to repeat. It requires money, coordination, time, and shared access to data. Even when a scientific community has the will and resources, the underlying measurements may not be stable enough to support the same kind of model training.
Most scientific data is messier
Protein crystallography, the key experimental technique behind much of the protein-structure data, is unusually reliable. The source article notes that over 25 Nobel Prizes have relied on it. In many other fields, experiments do not produce results with the same consistency.
Biology and chemistry often involve variables that are difficult to standardize. Cell lines drift. Chemicals can contain trace contaminants. Humidity in a lab can change. These details matter because modern neural networks need data that is consistent, accurate, precise, and scalable enough to learn from.
Some fields may still be ready for AlphaFold-style progress. The source identifies weather forecasting, much of genomics, and very limited areas of chemistry as places where the data conditions may already be close enough. It also says government support for producing and coordinating these datasets will be critical, as the US National Security Commission on Emerging Biotechnology has argued.
For most open scientific questions, however, waiting for perfect datasets is not a practical short-term plan. Science needs AI systems that can operate under uncertainty, because that is how human researchers already work.
AI agents fit the way research actually happens
Scientists rarely rely on one method in isolation. A biologist searching for new drug targets may compare docking calculations, known structures, molecular dynamics, and binding assays, then use judgment to weigh the limits of each result. The work is iterative: form a hypothesis, test it, revise it, and repeat.
AI agents are built for that kind of loop. In simple terms, an agent is an AI reasoning engine connected to tools it can use. Those tools can be digital or physical. Large language models have helped make these systems more flexible, reducing the need for narrow, scientifically specialized datasets in every case.
This makes agents different from tools that solve one defined problem. AlphaFold applies a powerful method to a specific class of questions. Agents are more general. Their value is that they can model parts of the human discovery process: reading, proposing, checking, ranking, refining, and learning from new evidence.
Google’s AI Co-Scientist is one example from the source. Announced in May, it was given a one-page brief and asked to figure out how antibiotic resistance spreads between bacterial species, an important driver of drug-resistant infections.
The system used sub-agents to divide the work:
- One generated hypotheses from the literature.
- Another challenged those hypotheses like a peer reviewer.
- A third ranked the strongest candidates through tournaments.
- A fourth improved the leading hypothesis.
The agent concluded that resistance genes were traveling on bacterial viruses, using whichever virus could carry them into a new host. According to the source, the hypothesis was correct. Researchers at Imperial College London had spent a decade reaching the same conclusion through wet-lab work, and their paper had not been seen by Co-Scientist because it was still in peer review.
The obstacles are real
AI agents are not mature replacements for scientists. The source is clear that they still face major technical problems. They can hallucinate, their judgment is inconsistent, and limits in memory and input size restrict how long they can operate on their own.
Those weaknesses matter in science, where a confident but false answer can waste time or mislead a research program. Any useful agentic AI system must be judged by whether it improves the reliability and clarity of research, not only whether it produces plausible ideas quickly.
Still, the source argues that these barriers will decline. If they do, the effect could compound across scientific work, especially in three areas: reproducibility, memory, and speed.
Why agents could change the pace of discovery
Reproducibility is one of the most important potential gains. Scientists have long been asked to share raw data and exact code so others can replicate their work. But that administrative work often arrives after the most interesting research is finished, and researchers have resisted it.
Agents could change that by automatically logging what they do. If every step is recorded as part of the workflow, other researchers can inspect and repeat the method more precisely.
Agents could also strengthen scientific memory. Knowledge in labs often moves through long training, observation, and old notebooks. As agents take on more of the research process, a lab’s history could become a central, standardized repository of institutional knowledge.
The largest change may be speed. The source describes an agent that can read a thousand papers in an hour, design 500 molecules, and learn from failed tests by morning. When testing becomes faster and cheaper, researchers can spend less time debating an idea and more time checking it.
That does not mean science is near its end. The article places agentic AI alongside broad tools such as calculus, statistical inference, spectroscopy, and the computer: technologies that did not close science down, but opened new problems. If agents become reliable enough, they may do the same across many fields at once.