Google DeepMind pivots to agentic AI for scientific discovery

At the Google I/O conference, DeepMind CEO Demis Hassabis announced a strategic pivot in the company’s approach to scientific artificial intelligence, moving from highly specialised tools towards agentic, large language model-based systems. The announcement coincided with the launch of the Gemini for Science package, which includes the AI Co-Scientist for hypothesis generation and AlphaEvolve for algorithm optimisation. This shift marks a departure from the era defined by single-purpose breakthroughs such as AlphaFold, aligning instead with a broader industry trend where general-purpose models are beginning to make independent contributions to fields such as mathematics and science.

The strategic realignment is evident in the reassignment of key personnel within the company. John Jumper, the Nobel laureate who led the development of AlphaFold, has reportedly moved from science-specific tool development to working on AI coding projects. This move is linked to Google’s efforts to improve its coding tools to compete with rivals such as OpenAI and Anthropic, but it also signals a prioritisation of agentic science, as coding capabilities are essential for the success of autonomous AI systems.

Google’s new direction is underpinned by recent developments in general-purpose reasoning models. OpenAI recently announced that a model described as being in the vein of GPT-5.5 disproved an important mathematics conjecture, demonstrating that non-specialised systems can achieve significant research outcomes. Pushmeet Kohli, Google Cloud’s chief scientist, published a piece in the journal Daedalus stating that the industry is moving toward AI that “begins to do science” rather than just facilitating it, a sentiment echoed by Hassabis who described the current moment as “standing in the foothills of the singularity.”

The Gemini for Science package aims to operationalise this vision by providing researchers with access to these agentic systems. While the tools are not yet publicly available, Google has opened applications for access, anticipating wider adoption within the scientific community. Early testers have expressed enthusiasm for the potential of these systems; Gary Peltz, a Stanford geneticist, described using the AI Co-Scientist as akin to “consulting the oracle of Delphi.”

Despite the shift in focus, Google is not abandoning its work on specialised AI tools. AlphaGenome and AlphaEarth Foundations, trained for genetics and Earth science respectively, were released last summer, and the newest version of WeatherNext came out in November. Isomorphic Labs, a Google subsidiary that utilises AlphaFold technologies for drug development, recently raised a $2 billion Series B funding round. However, the reallocation of resources and the emphasis on general-purpose models suggest a long-term strategy where agentic systems act as accelerants for human scientists, potentially evolving into collaborators in the coming decade.