Google’s bold agentic era aims to turn AI into an amplifier of human ingenuity. At Google I/O 2026, the company showcased how its research efforts are coalescing into a powerful suite of tools designed to accelerate scientific discovery and redefine the developer experience. Yossi Matias, Vice President and GM of Google Research, framed this year’s announcements around a central thesis: models are becoming more powerful, and the platform is becoming more agentic. This shift promises to make Google products substantially more helpful while transforming how researchers tackle the most pressing scientific and societal challenges.
The Agentic Era Vision
The overarching theme of Google I/O 2026 was the transition from static AI interactions to dynamic, agentic systems. Instead of simply responding to prompts, AI agents can now plan, execute multi-step workflows, and interact with the digital world on behalf of users. This represents a fundamental shift in how technology can serve as an amplifier of human capabilities. Google Research is embedding this agentic philosophy directly into its product ecosystem, from coding assistants to scientific research platforms. The goal is to create systems that don’t just answer questions but actively help users achieve complex objectives, whether that means writing software or curing a disease.
1. The Agentic Era Vision
This vision acts as the foundational layer for all other innovations. It repositions AI from a passive tool into an active collaborator. By focusing on agentic workflows, Google is enabling systems that can reason, plan, and execute tasks with a high degree of autonomy. This directly addresses the bottleneck in modern scientific research: the sheer volume of data and complexity of analysis required to make breakthroughs. The agentic era is not just about smarter models, but about more capable systems that amplify human intent.
Gemini for Science
One of the most significant announcements was Gemini for Science, a suite of experimental tools built on foundational research published in Nature. This suite is designed to expand the scale and precision of scientific exploration. It directly addresses the bottleneck in modern research: the vast gap between data generation and actionable insight. By integrating tools like ERA and Co-Scientist, Gemini for Science provides a unified platform where researchers can move seamlessly from hypothesis generation to computational experimentation. This represents a concrete step toward automating the scientific method itself, allowing human scientists to focus on creativity and strategic direction.
2. Gemini for Science
This suite represents a cohesive effort to bring the power of agentic AI to the global scientific community. It is built on the principle that AI can accelerate every step of the scientific method. The suite includes several specialized tools, each targeting a specific phase of research. By packaging these tools together, Google aims to lower the barrier to entry for advanced computational research, enabling smaller labs and institutions to participate in cutting-edge discovery. The google research io highlights clearly pointed to this suite as a cornerstone of their scientific strategy.
Empirical Research Assistance (ERA)
Empirical Research Assistance, or ERA, is a research coding system developed to help scientists write expert-level empirical software. Given a well-defined problem and a scoring system, ERA acts as a code-optimizing research engine. It proposes new concepts, writes the corresponding code, and evaluates the results. Through an iterative process of tree search, it explores thousands of code variants to optimize performance. The real-world impact is already visible: ERA has helped accelerate discoveries including predicting hospital admissions for respiratory illnesses and forecasting seasonal runoff across California’s river basins. These results signal the power of AI to unlock deeper insights through compute-driven experimentation.
3. Empirical Research Assistance (ERA)
ERA is specifically designed to overcome the challenge of writing high-performance empirical software, which is often a bottleneck in computational science. Its tree-search methodology allows it to explore a vast space of possible code implementations far more efficiently than a human could manually. This system has already demonstrated its utility in diverse fields, from neuroscience to cosmology. The ability to automatically generate and test thousands of code variants means that researchers can rapidly converge on optimal solutions for complex data analysis tasks.
Co-Scientist
Co-Scientist is a multi-agent system based on Gemini that functions as a collaborative AI partner for researchers. Rather than a single monolithic model, it uses a coalition of specialized agents that iteratively generate, evaluate, and refine hypotheses. This structure mimics the collaborative dynamics of a research lab, where different experts debate and build upon each other’s ideas. Researchers are already using Co-Scientist to tackle some of the most pressing scientific challenges, including antimicrobial resistance, plant immunity, and liver fibrosis. The system’s ability to propose novel, testable hypotheses marks a significant leap forward in AI-assisted discovery.
4. Co-Scientist
The multi-agent architecture of Co-Scientist is what sets it apart from standard question-answering systems. It hosts a “coalition” of agents, each with a specialized role, such as generating hypotheses, evaluating their feasibility, or designing experiments to test them. This internal debate and refinement process leads to more robust and creative scientific proposals. By acting as a true collaborator, Co-Scientist helps researchers break out of their own cognitive biases and explore novel avenues of inquiry that they might have otherwise overlooked.
Computational Discovery
Within the Gemini for Science suite, Computational Discovery stands out as an agentic research engine built with ERA and AlphaEvolve. This tool is designed to massively parallelize the experimental process. Instead of a scientist manually testing one hypothesis at a time, Computational Discovery generates and scores thousands of code variations simultaneously. This allows researchers to rapidly explore a vast space of modeling approaches that would take months to investigate manually. It effectively turns the computer into an automated experimentalist, accelerating the cycle of hypothesis formulation, testing, and refinement.
5. Computational Discovery
This tool is built for scale. By leveraging the code-generation capabilities of ERA and the evolutionary search algorithms of AlphaEvolve, it can automate the entire loop of computational experimentation. A scientist can define a problem and a scoring metric, and the system will autonomously generate, run, and evaluate millions of potential solution strategies. This capability is particularly powerful in fields like drug discovery and materials science, where the search space for viable compounds or structures is astronomically large.
Hypothesis Generation
With millions of scientific papers published annually, synthesizing existing literature to generate novel hypotheses has become a monumental challenge. Hypothesis Generation, another tool in the Gemini for Science suite, aims to bridge this gap. Built using Co-Scientist, it collaborates with scientists to define a research challenge. It then runs a multi-agent “idea tournament” where specialized agents generate, debate, and evaluate competing hypotheses. This process ensures scientific rigor by simulating the peer review and debate process within the AI system itself, helping researchers identify the most promising and novel directions for their work.
You may also enjoy reading: iPhone 18: Everything We Know Apple’s Most Ambitious Lineup.
6. Hypothesis Generation
The “idea tournament” mechanism is a core feature of this tool. It formalizes the creative process of scientific brainstorming by forcing competing hypotheses to be rigorously evaluated against each other. This helps to filter out weak or obvious ideas and surfaces high-risk, high-reward possibilities that might emerge from combining disparate fields of knowledge. It acts as a powerful engine for scientific creativity, ensuring that the starting point for any research project is as strong and innovative as possible.
Literature Insights and Agentic Peer Review
Staying current with scientific literature is a Sisyphean task. To address this, Gemini for Science includes Literature Insights, built with NotebookLM. This tool allows researchers to interact with complex scientific papers conversationally, asking questions and synthesizing information across multiple documents. Furthermore, Google is piloting a Paper Assistant Tool (PAT) for agentic peer review. This tool is being explored in collaboration with major conferences like ICML, STOC, and NeurIPS. PAT aims to assist the peer review process by automating routine checks and helping reviewers focus on the scientific substance of a paper.
7. Literature Insights and Agentic Peer Review
This final innovation tackles the critical but often overlooked aspect of scientific communication. Literature Insights makes the vast corpus of existing knowledge more accessible and actionable. Meanwhile, the Paper Assistant Tool (PAT) addresses the sustainability of the peer review system itself. By automating tasks like code checking and reproducibility verification, PAT frees up human reviewers to focus on the scientific merit and impact of a submission. This dual focus on consumption and quality assurance rounds out a comprehensive vision for an AI-enhanced research ecosystem.
Frequently Asked Questions
How does the Co-Scientist system differ from standard AI research tools?
Unlike standard AI tools that provide direct answers or summaries, Co-Scientist operates as a multi-agent system. It uses a coalition of specialized agents that iteratively generate, evaluate, and refine hypotheses through a process of debate and collaboration. This structure is designed to mimic the dynamics of a human research team, making it a proactive partner in scientific discovery rather than a passive information retrieval tool.
What makes the Empirical Research Assistance (ERA) system unique for scientific coding?
ERA is unique because it combines code generation with an empirical optimization loop. It doesn’t just write code based on a prompt; it proposes concepts, writes code, evaluates the results against a scoring system, and iteratively searches through thousands of variants using tree search. This allows it to discover novel and high-performing empirical software that a human might not have considered.
How can researchers gain access to the Gemini for Science tools?
Google is gradually opening access to the Gemini for Science suite through partnerships and pilot programs. Researchers interested in using tools like Computational Discovery or Hypothesis Generation are encouraged to engage with Google Research’s publications and announcements. The Paper Assistant Tool (PAT) is being piloted directly with conferences like ICML, STOC, and NeurIPS, indicating a phased rollout focused on the research community.
These seven innovations represent a cohesive strategy to embed agentic AI into the core of scientific inquiry and software development. By focusing on practical tools that augment human creativity and rigor, Google Research is laying the groundwork for a new era of accelerated discovery. The google research io highlights from this year clearly point to a future where AI acts as a true amplifier of human potential, turning ambitious scientific questions into tangible, testable realities.
