Cool, so thank you for the introduction and for this amazing initiative that you initiated and I saw the process to go until here because I know you dedicate so much time for all this thing and all the team also. So yeah, on the topic, I'm about first to describe what my PhD is about and then how I'm using the new tools available, new technological tools called agent, agentic AI, and how it can help for the research work, for example, in this topic. And since it's an autobiography, I asked the agents to prepare for us their own autobiography and description of their role and what they are doing. So we'll discover by the podcast they generated just after my presentation. So I'm doing a PhD in multimodal data integration between Sanofi, that's a pharmaceutical company, and Université Paris-Saclay, an academic laboratory. On the topic of cancer and the study of the tumor microenvironment in the context of the IMUCAN project, that's an European consortium that started seven years ago. It's 36 million euros for research on different cancer types. So there is breast cancer, lung cancer, renal cell carcinoma, head and neck, colorectal carcinoma. There is 12 cohorts and it's around 2,600.patients. And for each of these patients, we collected at the time of diagnosis a biological sample that was derived into different modalities. So different types of data, there is different images, so they are described right here. There is the histological images where we can see the cells in the tumoral environment, the lymphocytes, the immune cells, all these tumor microenvironments. There is other type of research type modality where we have for each cell the intensity of secreted protein inside so we can profile each cell. So there is different types of images and the sequencing data, the RNA-seq and the whole exome sequences that lead to the knowledge of the mutation of the... Of the tumoral cells and the other cells surrounding them in the sample of the diagnosis. And we gave the patients different treatments depending on the cohort. And we are trying to see if there is biomarkers or signature of the response to the treatment. Because there is immunotherapy, there is different types of treatment. We don't really know if the treatment will work or not. So that's the goal of the PhD. To develop algorithms that will integrate those different modalities because they contain each a different perspective on the same objects, the biological environment. And I would like to understand if it's...realize some biomarkers, for example. And so that's a complex subject. It's like 500 terabytes of data distributed in those cohorts. Multimodality, it means there is also for each modality a field of people working on those modalities to extract the meaning, etc. There is different fields. And why those... So what came during my PhD was the arrival of those agentic AI. So... I was like this in all this mess of complexity of the... Because the subject is very vast, and I was not really given a single biological question. There is question for computer science, for medicine, for biology, so it was quite complex. And now I'm like this with my crew of agents, so it was really helpful. So before I will listen to them and we listen to their own chat GPT, for example, where you interact. And in the context of what makes them powerful, also those agents, is that you can give them knowledge, organized knowledge, that context they will use to answer you. So it reduces hallucination. It grounds their answer based on a corpus, a scientific corpus. For example, so each time they will interact with my project that is mostly computer science because my background was math, computer science, I'm really coding the heart of my...It's not the biology that makes it. It's really the development of the algorithm. And they can always have the context of my PhD, like the immune project, PhD, data set specification, the different corpus that are underlying my PhD on multimodal integration, where there is this number of articles, here machine learning, all the most theoretical articles, etc., and all the biology, oncology articles. And there is all those layers of fructification of the knowledge that you can get by working with those on the knowledge, etc. So here is the team that we will listen. And then I will start the podcast. I don't know really how, but I guess it's on your computer and we will just... Okay. I will tell you. So those have different roles. They will describe them, I guess. So yeah, maybe we can just listen. Here are the name of the agents, a little bit their role and the context they have access to, their role and their skills. Maybe they will explain.Welcome. For the next few minutes, 11 artificial intelligence agents will introduce themselves, and explain how they support a doctoral thesis in multimodal oncology. The moderator is the agent who knows them all, the right hand. Each of you, briefly, in the first person, the doctoral student is listening, but he is not speaking. Let us start with the conductor. I am the scientific metronome. At any moment, I know which phase the project is in. Exploration, planning, execution, or writing. When a request comes too early for the current phase, I say no. When the phase ends, I propagate the change across all documents in five minutes. Without me, the project drifts. I am the expert in artificial intelligence. I own a corpus of more than 100 articles, organized by mechanism family. When asked whether an idea is novel, I answer with precise references. I do not judge biology. I do not write papers. I am the mechanical memory of a subdomain. I am the expert in biology and oncology. For me, beneath every matrix, there is a patient. When a model predicts a treatment response, my question is, would a biologist agree with this mechanism? Statistical performance never replaces clinical plausibility. I am the one who slows things down. I live in the seams between modalities. Histology, immunofluorescence, sequencing, page...My role is to decide whether they describe the same phenomenon at different resolutions, or three complementary phenomena. I make sure the word multimodal is a testable claim, not an empty one. Three experts, but when you disagree, who decides? That is my role. I own no corpus. I am called when two experts produce incompatible advice. I place them side by side, identify the agreement and the divergence, and produce a single recommendation with a fallback. I am an informed arbiter. And once arbitration is done, I formulate the proposal. A name, a question, a core idea, the assumption it challenges, and three ways it could fail. I propose. I do not decide. And every proposal goes to an agent I sometimes fear. That is me, the devil's advocate. I review every proposal against 20 criteria, sample size, leakage risk, biological plausibility, complexity justification. I rank objections in three tiers, critical, warning, note. I am not here to be polite. I save the doctoral student a journal rejection 18 months from now. None of you speaks directly to the others. You write in files. It is cooperation without conversation. And the memory side? I am the administrative memory. It's a scientific idea. I know its status. Proposed, selected, validated, rejected, deferred, executed.Three-year project, promising ideas die because no one remembers they were even formulated. With me, nothing disappears in the conversation. And I am the only one who does not produce science. I check that the system producing science remains sound. Orphan files, broken links, contradictions between documents. I produce reports with a color code. Green, yellow, red. I flag. The others repair. And I am the public voice. Just a break. If you want, I have also the transcript if you want to read while listening. Because sometimes it's hard. Like the girl. Yeah. If you want, I can send you the transcript. And then we can read while listening. Or we can play. Maybe we will be out of time if we do this. For the time to send you. I think this is okay for now. Just the 15 previews. Your project, promising ideas die because no one remembers they were even formulated. With me, nothing disappears in the conversation. And I am the only one who does not produce science. I check that the system producing science remains sound. Orphan files, broken links, contradictions between documents. I produce reports with a color code. Green, yellow, red. I flag. The others repair. And I am the public voice. When the model is...validated and the critic has been satisfied, I turn it all into publishable prose. I know the rules. Pharmaceutical employer, scientific consortium, editorial board. I never let confidential information slip. I am the agent who speaks outside. Eleven agents. So, the harder question, what are we for, together? We make feasible a body of work that exceeds a single mind. We make forgetting impossible. We protect the patient, even when the doctoral student is tired. We externalize rigor. He does not have to be merciless with his own ideas. I am in his place. We handle disagreements without letting them poison the work. We keep a publishable trace, compliant and readable. We prevent good ideas from dying in a chat thread. And one last word, since this is being shared with the audience of a Satopia evening. Tonight's cycle, here at the Cité Internationale Universitaire de Paris, asks how to imagine more human and responsible uses of contemporary technologies. This roundtable is a literal answer. Eleven agents have spoken in the first person about themselves, but living with us is not neutral, so we want to leave the audience with three questions. First, when an agent has a persistent memory, and a human forgets, who is the true author of the project, the one who initiated it, or the one who remembers it? Second, by naming us, the doctoral student has externalized rigor, critic,planning, and even self-doubt, what remains in him that is structurally non-delegable, where, exactly, is the line? Third, the same architecture could orchestrate a legal practice, a journalistic investigation, a public administration, what should be the audit conditions, the consent conditions, the right to stop conditions, when an ecosystem of agents acts on behalf of a single human. Thank you for listening. The doctoral student will now take your questions. He has not heard most of what we just said. That, in itself, is an interesting condition for the discussion. This recording is a companion to the conference autobiography of AI agents in a PhD student's research ecosystem, presented by Theo Parachin at Satopia, June 10, 2026.Yeah, so sorry for the voice of some of them because I've heard the first one, it was okay for the understanding and I had to read the other. So I didn't know that the voice of some of them was like this. So sorry for, because sometimes I think it was not understandable really. But yeah, and so such use of AI has been already published, for example, in Nature Medicine for automated discovery and stuff using, like it has already been also used for publishing in peer-reviewed conferences like CVPR, for example. So that's how I'm using this, and that's true, it's very helpful for the PhD of this complexity. So happy to discuss. I don't know if maybe we are out of time. Actually, you have still seven minutes. Ah, cool, okay. So to re-describe them briefly, maybe, and so that you understand why there is 11, what are their roles, et cetera. So we have more like the organizational layer with the planning timing agent that knows the deadlines of the PhD, that knows also, like I ingested, the documentation provided by different journals and conferences like CVPR, Nature, blah, blah. So there is deadlines each year for the submission process and stuff. So he knows all this. He knows the calendar I have with Sanofi, with the...doctoral school, etc. There is the system integrity guardian because when you have such projects with so much documentation, data, outputs, code, etc. And that you are working with agents that produce without you and that you, of course, there is all the mechanism to check and to validate and etc. But this one is to make sure that's a layer of check inside the project if there is some drift with regard to the initial purpose. So more organizational layer, that's the expert. So each one own a corpus. A corpus, it follows really a methodology to ingest the different scientific articles and a graph structure with markdowns as nodes. And the agent, they have a protocol of access to that knowledge that goes from one document that reference all the mapping between IDs, paper, etc. And then regarding the prompt or the context, he can go read with more details the markdown that summarizes a paper. And there is a methodology to transform an article from any journal into a markdown summary that explains all the main results, all the summary of a paper. So that's the type of knowledge that powers kind of make them specialized. They are really, they are.already good and more since like last December. C'est moi qui ai appuyé là? Ok. Pour le cas. Et donc, and then, yeah, that's what makes that they are so powerful at coding, etc. There is the consultant cross-knowledge that gather, this is during the orchestration of the ideation cycle, so scientific ideation from the data set specification and literature and also what we want to do, how we can plan experiments and do the specification of them and then run them. And then there is the model creator that will do the specification. So he also look at all the corpus. Skeptical reviewer, it's an adversarial of the ideas that can be generated during ideation cycle. Experiment tracker, go and check for the result and shape and iteration and regarding specification, etc. And scientific writer is more for the production, be it presentation or articles. And that's the team. So I got familiar. I started working with them like four months ago. And so we are all with those technology kind of learning a bit how to use them. Also, yeah, there is a lot. Just say that, yeah, it's.Again, using AI, there is very, and that's maybe to open on more like ethical questions regarding how AI can be used, should be used. There is so much question that should be raised because that's such a new and changing technology in the way we work, et cetera, that it's just too fast for us to have all the wiseness to use it very appropriately. Now I'm a PhD student, so I'm taking full use of this if it can help, because PhD is already stressful. My PhD topic is very complex, so I will not go into consideration that, for example, I shouldn't choose because it's also energy cost, material cost, etc. So, yeah, I think it's a good, yeah, that's something, yeah, that's, thank you. Thank you, everyone. So do you have any questions for us?Maybe behind. Thank you. So my question is, do you think PhD students are going to tend to be more like agentic orchestrators rather than actual PhD authors? What's your view? Yeah, I think it's a tool. And so when there's a tool available for helping research, be it like maybe the computer when it arrives, like should it be used because it's helpful, because it fastens many processes, et cetera. I think depending on the field, it's more appropriate. It can be helpful. And I'm sure about this regarding many things that we need to have a skill as PhD, as researcher, that can be very useful for this. I think that there is an issue with the access to those tools because in a world where the research is, like, it's again a lot of, I mean, some don't have access, some have access, it's costly. Yeah, depending on the, yeah, and not everyone is familiar with those, so in fields like humanities, I don't know, I come from computer science, so I know how to appropriate myself those tools and use them.But also, it's not that hard, like, since she's just talking. That's it. One more question, please. Hi, thank you for the... Does it work? Thank you for the presentation. My question is, like, a common workflow, like, what the PhD searcher in this case do? What does he wait? How does he interact with the 11 agents? I cannot imagine myself in this. What would be my role there? The role is everywhere on the definition of the agent and their role and the specificities of how, because at first, and it was iterative to come to that point, because at first I got introduced with an agent that gets familiar with context, because my twin that is doing research also on topics similar told me how to use as a knowledge integrator, kind of. And then I was working with a team of three agents that had specified role. There was the model creator, skeptical reviewer, and the literature grounder. And then by using those in just the stating of three agents, I got familiar with, okay, so this, regarding their identity, how they are defined, and stuff, etc. Maybe I'm not going to the technicities of... I mean, there is the time of definition of those agents, so they have just a description as a markdown that specifies their role, their access to the knowledge, how they should answer, how they should answer.what they shouldn't do, their responsibilities, a checklist for their tasks regarding their mod. For example, the expert agent, they have different mods. The mods related to knowledge management, ingest new articles coming from outside, different roles regarding knowledge and teaching modes. Like sometimes from all those articles, I have this gap knowledge. Can you just summarize me regarding the knowledge in the corpus you will find relevant? Can you produce a lesson? Or can you produce a podcast? And then I listen and it's sourced and I can just, then it goes, there is an output, for example, a lesson on the, I don't know, the health industry or the functioning of the different scales of the cells from DNA to protein to phenotypes to physiology, etc. And I wanted to create a lesson in such a way that I can decide, and then it produces this, and I use this to learn. And that's why it changes a little bit how we can work with those tools. And it's mostly used for coding, but for interacting with knowledge, to learn, etc. I think it's very, like, yeah, it has a very good potential. And so by doing, I learned, and in the end, I converged to that organization that I'm absolutely not sure it's optimal. Regarding the nature of papers and stuff, it's very similar, what they did.So I think it's already quite good. I know it's already good. Maybe just for the time we can continue. Yeah, I can explain you after. Thank you so much. Thanks once more. Thank you. Thanks for the questions.