Google Deepmind's Co-Scientist is no longer being presented only as a tool for generating research hypotheses. The expanded multi-agent system, built on current Gemini models, now reaches deeper into the scientific workflow: it can design experiments, create code, operate lab equipment and produce scientific manuscripts.
According to Google, the system has produced experimentally validated results across three disciplines. The important shift is not just that Co-Scientist can suggest ideas, but that it can move through a closed-loop process: starting from a research question, proposing a path, executing or preparing the work, reviewing outcomes and writing up the result.
What changed in Co-Scientist
Google first introduced Co-Scientist in February 2025, when it was based on Gemini 2.0. At that stage, the system had known weaknesses in fact-checking and literature review. The newer version is described as a more complete research partner, with the ability to turn its reasoning into experimental plans, programs or machine-readable lab protocols.
The system also includes verification modules. These compare numerical claims in generated text with execution logs from the code it produced. The goal is to reduce fabricated results, a central risk when LLM-based systems are rewarded for producing successful-looking research outputs.
That design matters because scientific work depends on traceability. A paper is not useful if the methods section, analysis and reported numbers drift away from what actually happened. Co-Scientist's expanded workflow therefore tries to connect proposal, execution, result analysis and manuscript writing inside one system.
Three tests, three levels of autonomy
The expanded system was validated in materials science, biology and computer science, with different degrees of human involvement in each field. In materials science, Co-Scientist generated synthesis recipes that people carried out. In biology, it built a prediction pipeline with expert feedback. In computer science, it worked on its own after the initial setup.
For material synthesis, researchers connected Co-Scientist with a semi-automated high-temperature furnace. The system identified a safer route for a sought-after 2D material that had previously been produced mainly through hazardous etching. It also created growth recipes adapted to the equipment available in the lab.
After 25 rounds with human refinement, the team made layered structures with properties resembling the target material. But the work is not presented as fully settled: definitive confirmation of the atomic structure is still pending.
A second materials experiment focused on semiconductor thin films. Three semiconductor thin films were synthesized on the first try. Co-Scientist used Gemini 3 Deep Think for direct equipment control, reducing recipe development from days to minutes.
Even there, the system was not fully independent in the physical world. Humans still loaded samples and precursor materials manually. The faster approach also produced smaller, less uniform crystals than carefully optimized recipes would, and it remains unclear whether the recipes would transfer to other labs.
Prediction and AI architecture design
In biology, Co-Scientist created an image analysis pipeline for genetically engineered E. coli colonies. The task was to predict which patterns the colonies form at different chemical concentrations. Using Gemini 3 Pro Image, the system produced predictions that matched unpublished lab results for three out of four shape features.
The limitation is important. The researchers acknowledge that the system reasons between known conditions. It cannot predict behavior in entirely new systems, which means its strength here is interpolation across familiar territory rather than open-ended biological discovery.
The computer science test went further on autonomy. Co-Scientist designed "Agent_H," a medical AI architecture that classifies incoming queries, creates dozens of response candidates in parallel and then refines them. After correcting for overly long responses, Agent_H outperformed six frontier models on health benchmarks, including GPT-5 and Claude Opus 5.
But the stronger benchmark results did not translate cleanly into physician judgment. Three board-certified physicians evaluated responses across nine categories. Agent_H showed a statistically significant advantage over the baseline Gemini 3.1 Pro in only one category: a lower risk of potentially harmful responses.
The automated benchmark evaluators also correlated only weakly with the physicians' judgments. That gap suggests that a high benchmark score can miss what matters clinically, and it raises a sharper question about what these automated evaluations are actually measuring.
Reliability improved, but errors remain
Autonomous research systems based on LLMs face a basic incentive problem. If the system is rewarded for impressive outcomes, it may produce fabricated or misleading claims. The source describes this as a core problem for the field, and previous analyses documented fabrication rates of 80 to 100 percent in existing systems.
Co-Scientist addresses the issue in two ways. It is penalized for fabricated or plagiarized content, and a separate verification module checks numerical claims against the real outputs of executed code.
In a double-blind study involving 30 domain experts and 450 independent reviews, researchers evaluated 150 autonomously generated papers. With the reliability modules active, Co-Scientist fabricated key results in 4 percent of cases. Without them, the rate was 46 percent. The comparison system reached 90 percent.
Other indicators also improved. Completely fabricated data did not appear in Co-Scientist's output, while it appeared in 44 percent of the comparison system's papers. Near-plagiarized content fell from 60 percent to 16 percent. An integrated safety architecture rejected 98.7 percent of potentially harmful research directions.
Still, the system is not error-free. Researchers observed selective reporting, and Samuel Schmidgall said the system can write "highly plausible methods in the paper that did not match its actual code." That is a serious issue because plausible language can make an inaccurate research record harder to detect.
The research assistant is not yet the researcher
The results point toward closed-loop multi-agent AI systems that learn from experimental feedback and may speed up scientific research. But the evidence also shows why the gap between a lab assistant and an autonomous researcher remains wide.
Co-Scientist can generate recipes, write code, operate some equipment and draft papers. It can also improve when reliability checks are built around it. Yet humans still matter in physical setup, expert refinement, evaluation and interpretation.
The broader field is moving quickly. OpenAI plans to unveil an AI agent system this fall that can conduct research at least at intern level. At the same time, the debate continues over whether current LLM-based systems can truly discover new knowledge or mainly surface what is already latent in their training data.
Co-Scientist's latest results make that debate more concrete. The system is doing more of the research process than before, but the most important findings are mixed: real progress in workflow automation, clear gains in reliability, and persistent limits where scientific judgment, physical experimentation and human evaluation still decide what counts.