From Research Assistant to Discoverer: Are AI Agents Becoming Scientists?

Hello, Victoria. Can you hear me? Yeah, I can hear you. Hi. Okay. Okay, so we have two guesses haven't arrived yet. But just let us get started. Okay. Hello, everyone. And welcome to today's joint care space. Our topic is from research assistants to discover AI agents becoming scientists. We'll be exploring the intersection of AI agents, scientific discovery and decentralized science. Today's discussion was inspired by a recent anthropocresearch program in which roughly 955 cloud agents spent 21 hours analyzing genomic data and scientific literature. The system, surface of previously uncactorized enzyme system, associated with DNA repeats arithesis, remedy sense of whisper. Human researchers then examined and validated the findings while the biological function and the practical value of the system still require further study. Dix example raises a fascinating question. If AI agents can search enormous scientific spaces, identify patents and generate promising research leads. When does their work become more than assistance? And when, if ever, can we call it a generic discovery? We also discussed how humans and the AI should divide the work. How researchers can evaluate AI generated findings? And whether this model could create new opportunities for smaller teams and the design communities. We are joined by four guesses working across AI, Web3, scientific infrastructure and the design. Thank you all for being here. Okay, so let's begin with a quick introduction for each of our guests. Please tell us who you are and what your project is working on. So first, let me invite Natasha from JoinCare. Could you briefly introduce yourself and join Care? Hi, first of all, I'd like to say my internet connection is a little bit suspect. So if I cut out at any time, please tell me immediately. So I'm Natasha and I am the CTO of JoinCare. I am a machine learning engineer. I'm also an award-winning founder and entrepreneur. And I have a highly interdisciplinary background that includes mathematics, philosophy and artificial intelligence. My experience spans the AI pipeline from data acquisition and analysis to custom model building and the post-training and evaluation of larger systems. At JoinCare, I focus on the development and deployment of our care agents and on how AI can be applied rigorously and effectively across project evaluation and research and development. Thank you, Natasha. Victoria, could you also briefly introduce yourself and also about your project? Thank you so much for having us. I'm Victoria. Community Manager at Brinkville. Brinkville is an on-chain liquidity and execution of the software for AI agents. So we build a collaborative liquidity pool and agent agent settlement rails. So the financial pump in that, like, autonomous AI systems, transact with each other without requiring a human to approve every single step. So the connection to today's topic is really direct because the research model describes in the anthropic scene. I think about 950 agents running in parallel for 21 hours, sufficient of previously uncharacterized enzyme system. I think it's not just an AI story anymore to us. I think it's more of an infrastructure story. So those agents, they actually needed a compute and they generated an output that required coordination and the story. And any version of this that involves rewarding contributors or funding experiments or distributing findings to broader research community, required financial set-manager structure. And RIMFIL is the infrastructure player that makes AI agents economic activity. So including scientific research, activity, programmable, transparent and on-chain. So that's what we do and that's why we are at this table today. So yeah, I'm pretty excited to hear from the rest of the speakers. And I learned more and was like the country with as much as I can. Thank you so much for having us. Okay, thank you all of our garrisons, introductions. So let's move into our question one about what can AI agents already do? So AI systems have moved far beyond the same poor question and answer tools. Agents can now search literature, analyze data sets, compare hypothesis, coordinate sub-taskers and surface patents that may be difficult for small human team to find manually. So the question is what can AI agents already do for scientists today? At what point can we call their output a generic scientific discovery? So let me invite Natasha. So where do you think AI agents already create 18 real value? Yes, so let's start with the first question here. What can AI agents already do for scientists today? And I would actually quickly like to talk first about what AI in general has already been doing for scientists for a while. And then specifically what AI agents have brought to the table. Because I think that distinction is quite important for your second question about what counts as genuine scientific discovery. Now, in science AI models, so not agents, just AI models, have already transformed a lot of the research process. In the life sciences in particular, systems such as alpha-fold three, which I think most people have heard of at this point, and diffusion models have completely mis-shaped protein structure prediction and molecular generation. AI is widely used for drug screening, multi-target drug design, and admet prediction, so predicting properties such as absorption, metabolism, and toxicity. It is changed by a logic and natural product discovery through things like peptide generation, antibody optimization, protein engineering, and virtual screening. So the models I have just described have already been hugely significant in healthcare and drug discovery. However, they are also fundamentally passive. They respond only when asked, and this is where AI agents have really taken things up and not. Unlike these former models, AI agents are autonomous and they are active, and this means that they can follow that continuous iterative workflow that usually underpins scientific discovery. So, more concretely, an agent can collect and perceive data from multiple sources, reason over and analyze it, decide what to do next, and then act. And in a bio-pharmaceutical context, that might mean recommending a molecular modification, choosing which candidate to test, or sending an experiment to an automated lab. Agents can then also compare their predictions with real-world experimental results, or expert corrections, and learn from that feedback to refine future decisions. So, AI agents really have the potential to handle an entire scientific workflow on their own. Not only that, they can work continuously, and at far greater speed than humans, and this is of course having a huge impact on the efficiency and cost of research. And to increase this efficiency even further, a lot of people are making use of multi-agent ecosystems, which means that instead of relying on a single AI agent, organizations that provide specialized agents that collaborate. For example, to identify drug candidates that need to meet multiple requirements simultaneously. Now, let's talk in more detail about some impressive feats that AI agents achieve across the scientific workflow. Firstly, as I touched on already, they can read and analyze research at a scale that would be impossible for an individual scientist. The multi-agent AI scientist Krosmus reads around 1,500 papers and executes about 42,000 lines of analysis code in a single run. The AI agent open-scroller searches across 45 million open access papers and produces citation-backed scientific synthesis. And more recently, there's actually been a quite inventive new framework called paper to agent that goes beyond just reading papers. It actually takes a paper's methods code and data, and builds an agent that can apply those methods to new research questions. Agents build from different papers can even work together. And in one demonstration, the system combined methods from several papers to investigate an ADHD-associated genetic signal, and generated a new mechanistic hypothesis for testing. And that brings us to the next important part. AI agents can generate, compare and refine hypotheses. For example, Google's co-scientist uses multiple agents to propose and rank hypotheses for experimental testing. There's also a multi-agent system called Robin that proposes hypotheses and experiments, that comprises the resulting laboratory data, and then generates revised hypotheses. Agents can also take over its really specialist computational workflows. So, for example, cryoagent autonomously coordinates end-to-end cryo electron microscopy image processing. And that includes tool selection, and it also includes recovery from failed processing steps. And finally, agents are increasingly able to design experimentally testable biological molecules. For example, Stanford's virtual lab used an AI principal investigator and specialist scientist agents to build a nano-body design pipeline and generate 92 new SARS-CoV-2 nano-bodies. But the capability frontier is moving from reading and analyzing science towards generating hypotheses, choosing experiments, interpreting results, and producing candidates for laboratory testing. And so, this brings us to the second part of your question, which was, at what point we can call this output a genuine scientific discovery? Now, as I was saying at the beginning, AI in general has been transformative in scientific discovery for a while now. But I think that in the past, we have thought of it as a tool, as a passive resource that scientists can use. And traditional accounts of scientific discovery have placed a lot of emphasis on the idea of agency. You know, the idea that there is a specific person or group of people who are responsible for conducting or, or at least the important part of a discovery process. But agents have that agency. It's even in the name. They are becoming more like scientists in their own right. And for me, if AI agents are making breakthroughs that are genuinely novel, significant for field and have been verified, I see no reason not to call that a scientific discovery. But this also creates a real paradigm shift in science. Because now that the digital role of the researcher is changing, we are suddenly able to ask a whole new set of questions about who scientific discovery belongs to. Because now, discovery might still involve the researcher who framed the question, but it might also involve the AI agent that generated the hypothesis, the developer behind it, the reviewers, or a lab that validated it, even the patients who provided the data or a community that financed the work. And this is where DSAI organizations like Joint Care will really be able to redefine how science is done in the future. And finally, one thing to think about is that whether something is a discovery and who gets credit for it are separate questions. Because was it finding is verified? I mean, it's a discovery no matter who or what made it, but how to credit it remains an open problem. And at dsi.billian in 2026, legal panels discussing autonomous science, emphasized that AI generated discoveries still have to fit within existing human-centered IP and ownership frameworks. And here, dsi can really help because timestamped records of which agent proposed what and when could help establish priority and make individual contributions much easier to trace. Okay, thank you, Nathasha. It's brilliant. And so let me invite victory. What do you think the question about what can AI agents already do? Victory? Yeah, I'm here. I can hear you. So I think the question of when AI after becomes a genuine scientific discovery, it's I think it's really important. So our part of the nation is not only our discovery, I think every search engine finds like pattern and every statistical model identified the collaboration. The bar for discovery is like really high right now. And I do agree with the last speaker. So I was thinking like a discovery is when you're finding a change that changes what the scientific community looks for next. So that's what I really think about when I think about discovery. So when it opens a new search space whether bread is in just like