Google I/O Elevates Agentic AI for Science While Retaining Domain‑Specific Models

Demis Hassabis closed Google I/O by saying the field is “standing in the foothills of the singularity,” and the company used the keynote to cast agentic, general‑purpose AI as the next phase of scientific research. The presentation included a DeepMind video showing WeatherNext, the company’s weather‑prediction software, which Google says provided an advance alert about Hurricane Melissa’s catastrophic landfall in Jamaica last year and potentially helped people avoid harm. That framing signals a potential reorientation of research workflows and funding toward systems that can drive investigations with less direct human steering.

Alongside the agentic rhetoric, Google emphasized concrete product rollouts and domain‑trained models. The newest version of WeatherNext was released in November; AlphaGenome and AlphaEarth Foundations, models trained for genetics and Earth‑science use cases, were released last summer; and Google reports that AlphaFold protein‑structure predictions have been used by more than three million researchers worldwide. These releases show the company is shipping tools intended for immediate scientific and operational use, not just conceptual demos.

The keynote set agentic, LLM‑based agents against specialist scientific tools and highlighted internal shifts to support that trajectory. Google is cultivating general‑purpose agents and has reassigned contributors — including AlphaFold co‑creator John Jumper — toward coding and agentic work. The company also uses cautious product names, labeling one system “AI Co‑Scientist” rather than “AI Scientist,” even as both internal and external narratives push toward increasingly autonomous scientific agents.

That tension echoes broader changes across the field. Proponents argue agentic models can carry out research with limited human guidance; recent examples include an OpenAI general‑reasoning model that reportedly disproved a significant mathematical conjecture. Academics and Google Cloud’s Pushmeet Kohli have written explicitly that scientific work may shift from AI as a facilitation tool to AI‑led discovery — a framing that influences where funding and engineering resources are directed.

At the same time, Google’s messaging and activity make clear it is not abandoning specialized systems. Continued product launches and updates, the extensive use of AlphaFold outputs, and the recent $2 billion Series B for Isomorphic Labs-an entity leveraging AlphaFold‑related technology for drug discovery — indicate sustained commercial and research demand for targeted, domain‑trained models alongside general agents.

For builders, researchers and engineering teams the implications are practical and immediate: expect platforms and APIs to support both agentic orchestration and domain‑specific models, and prepare for integration challenges such as experiment tracking, validation, and human oversight. The specific rollout dates and adoption figures cited at I/O underline that this is an active engineering trade‑off shaping toolchains and priorities today, not a distant abstract debate.