Ever since Google DeepMind’s AlphaFold cracked the half-century-old protein-folding problem in 2021, AI’s role in science has usually been characterized in terms of the quest for similar big breakthroughs—proof that machines could solve problems humans could not. Anthropic, however, is pushing a different idea: that AI agents may matter more in the unglamorous work between discoveries.
In exclusive interviews announcing new partnerships with the Allen Institute and the Howard Hughes Medical Institute, Anthropic’s head of life sciences Jonah Cool and Grace Huynh, executive director of AI applications at the Allen Institute, said the elite science labs are using Claude-powered AI agents to tackle the analysis, annotation, and coordination bottlenecks that can stretch research timelines into years.
A ‘compressed 21st century’
Cool, a cell biologist and geneticist by training as well as a technology leader told Fortune that he was inspired by a 2024 essay by Anthropic CEO Dario Amodei, Machines of Loving Grace, which argued that “AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50 to 100 years into five to 10 years.”
It’s an idea Amodei described as a “compressed 21st century” that could make possible everything from near-universal prevention of infectious disease and major reductions in cancer mortality to effective treatments for genetic disorders, Alzheimer’s, and other chronic illnesses. Amodei also suggested that AI could enable highly personalized therapies, expand human control over biology itself, and even dramatically extend healthy lifespan.
For Cool, that vision maps directly onto the use of AI agents in science—not as tools that deliver breakthroughs, but as systems that can take over time-consuming analysis, coordination, and experimentation tasks that slow discovery across labs, allowing humans to potentially make those critical new discoveries.
“What AlphaFold achieved is incredible,” said Cool, referring to the system’s solution of the long-standing protein-folding problem. “But what we’re talking about here is different. It’s about working with teams across the scientific process and embedding AI into their daily work.”
Huynh said that the move towards AI agents at the Allen Institute, a non-profit bioscience research organization founded in 2003 by Microsoft cofounder Paul Allen, builds on tools many researchers are already using, particularly Anthropic’s Claude Code, which has become popular among computational biologists. In addition, the goal, she said, isn’t to apply AI everywhere, but to focus on specific parts of the research process—such as data analysis tasks that can take months—where agents can have the most practical impact and meaningfully speed up scientific work.
No single researcher can see every connection
We’re starting to reach a point where ‘big science’ is the norm,” she said. Scientists generate so much data today—from single-cell genomics and massive imaging datasets to connectomics, the study of how neurons in the brain and nervous system are connected—that no single researcher can hold it all in their head or see every connection anymore.
Cool pointed to the Allen Institute and the Howard Hughes Medical Institute as ideal partners precisely because of the role they already play in shaping modern science. The Allen Institute has produced some of the world’s most widely used biological datasets, including detailed maps of the mouse brain that show where genes are active in actual tissue—resources that have become standard tools for researchers across fields, not just neuroscience. More recently, those maps have been pushed to single-cell resolution, dramatically increasing their scientific value while also making them far more complex to analyze.
And at HHMI’s Janelia Research Campus, researchers have developed foundational tools such as calcium indicators like GCaMP, which allow scientists to watch neurons fire in real time, and advances in super-resolution microscopy that helped push past the physical limits of light imaging. The emphasis on tools and datasets, Cool said, is exactly what makes these institutions fertile ground for AI agents: speeding up analysis, annotation, and coordination there doesn’t just help one lab—it ripples outward across science as a whole.
“Science is a fascinating but highly repetitive and often very tedious practice,” he explained. “Increasingly in science, what that means is a lot of work related to analysis and transformations of data sets,” he said. “I think we’re like approaching a world where that will still take a lot of work, but…you get to the next steps and the experiments much, much, faster.”
A future where AI can help make hypotheses
Cool also described a future in which AI agents don’t just analyze results, but help scientists decide which hypotheses to pursue—narrowing hundreds of possible experiments down to the few most worth running, and even proposing novel DNA designs based on patterns humans alone can’t easily see.
“We’re moving towards the models being able to help make hypotheses,” by building off of the knowledge humans have already, he said. “We’re starting with, ‘help me prioritize the hypotheses I have,’ because I have a limited amount of resources, and I want to do all 100 experiments, but I only have money for 10.”
