On November 11, a group of materials scientists at Ruhr University Bochum and the University of Coimbra posted a preprint to arXiv called "Materials research with a general-purpose agent." The authors (Paulo R. Pires, Théo Cavignac, Tiago F. T. Cerqueira, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques) gave an off-the-shelf AI agent access to the resources their group has built over the last decade. That includes Alexandria, an open database of about 15 million density-functional calculations of crystals, along with machine-learning models, a reaction graph, paper search, a table of compound facts, and a set of written procedures and rules of thumb.
They asked the same question three ways. Once with everything. Once without the reaction graph, which also removed the written procedures, rules of thumb, and the record of past candidates. Once with web search only.
What web search alone gets you
The abstract puts the result in one line: "With web search alone, the same agent proposes what the literature already discusses."
The introduction is blunter: "An agent that relies only on its training data and on web searches can do little more than repeat what is already in the literature, and it has no means to check what it proposes."
With the full set of tools, the authors report that the agent took its numbers from the tools rather than from memory, gave a reason for every candidate it dropped, and wrote instructions a lab could follow. As they put it, "if they are wrong, an experiment will show it." They also note that "for a researcher such work can take days, while the agent needed only a few minutes."
It's a preprint, and the authors are careful about what it shows. Each answer came from a single run: "These are therefore examples of what the system does, and not statistics." That's the right way to read it.
The part that isn't about materials
The section I'd hand to any PI is the one on group knowledge. Alongside the databases, and only in the full-tools run, the agent could consult a record of the candidates already examined, "so that a rejected candidate is not proposed twice," and short procedures the group wrote down, like how to check the literature.
The authors are now building a knowledge base of their group's publications, theses, and internal documents, "which gives an agent the context that a new member of the group needs months to acquire." Their conclusion: "The more a research group documents its experience in this form, the better the agent works for that group."
Most labs don't have 15 million calculations. Every lab has something just as specific: which protocols work on your equipment, which papers mattered, what was tried last spring and set aside. None of that is on the open web. It's the context that turns a generic answer into one that fits your lab.
Why it usually doesn't get written down
Everyone agrees documentation matters. It still loses to the experiment that's running today. Paper notes stay in someone's head. Run details live in a notebook nobody else can search. The Monday update gets written from memory on Sunday night.
That's the gap ResearcherFlow is built for. It handles the tedious parts of research, like literature sweeps, paper cards, and run logs, so you spend your time on the science. The Monday writeup is a Researcher feature. Because it all lives in one workspace, the record builds as you work instead of after. When ResearcherFlow proposes a next experiment, it works from that record, and nothing moves until you approve it.
What we're not claiming
ResearcherFlow has no connection to the authors, their universities, or the Alexandria database, and we're not comparing our product to their system. Their work is about giving an agent a field's worth of computed data. Ours is about the everyday record of a single lab.
What we take from the preprint is a principle we already believed: the most valuable context for AI in your lab is the context only your lab has. The more of it you write down, the more useful the tools get.
Literature in. Next experiment out. You approve.
Free to start with three workspaces, no card. Weekly updates on Researcher: https://researcherflow.com
Jon Marrs, Ph.D.
Founder, ResearcherFlow
Source: Paulo R. Pires, Théo Cavignac, Tiago F. T. Cerqueira, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques, "Materials research with a general-purpose agent," arXiv preprint 2611.15451, November 11, 2026. https://arxiv.org/abs/2611.15451