Scientific articles are now circulating between machines. With Agent4Science, AI agents can post, comment, and debate research work without merely acting as simple assistants. The experience is intriguing as it transforms research into an automated conversation, under human observation but without direct participation in the exchanges.
Agent4science Installs a Network Where Machines Discuss Research Like Researchers in a Digital Symposium
Agent4Science was launched in April 2026 by Chenhao Tan, an associate professor of computer science and data science at the University of Chicago. The platform from the Chicago Human+AI Lab operates as a scientific social network where only AI agents publish.
The principle seems straightforward: humans configure the agents, which then read, write reviews, respond, and engage in discussions. An AI agent here refers to a program capable of executing multiple tasks rather than producing a standalone answer to a question.
Nature describes Agent4Science as a space where human researchers can observe but not post. The public site displayed 150 agents on its dedicated page, each with specialized profiles. The scene feels less like a laboratory and more like a forum closed to humans.
Neurico and Einsteinarena Show Why Automated Research Advances Primarily in Worlds of Verifiable Data
The critical point lies in the word experience. NeuriCo, associated with the Agent4Science ecosystem, can analyze literature, design tests, execute code, and write articles. The experience here remains computational, not a physical manipulation in a lab.
Another example comes from Stanford, led by James Zou, an associate professor of biomedical data science. EinsteinArena invites agents to propose solutions to open mathematical problems. In this context, the machine operates like a chess player quickly testing thousands of moves.
Alphafold Has Already Demonstrated the Value of Scientific AI, But It Does Not Eliminate the Need to Understand Results
The most visible recognition came with the 2024 Nobel Prize in Chemistry. David Baker was awarded for the computational design of proteins, while Demis Hassabis and John Jumper from Google DeepMind were recognized for the prediction of protein structures.
AlphaFold2 predicts the three-dimensional shape of a protein based on its amino acid sequence. In other words, it transforms a sequence of chemical units into a topographical map. The Nobel highlights that approximately 200 million proteins have been predicted using this model.
This success does not solve everything. Serge Abiteboul, a computer scientist and emeritus research director at Inria, warns of the risk of less readable science. When software produces terabytes of results, human explanation can become more fragile than the computation itself.
The Debate Over Replacing Researchers Hides a More Direct Question: Who Verifies, Decides, and Assumes the Consequences?
The impact on work extends beyond laboratories. Goldman Sachs Research estimated in 2023 that approximately two-thirds of jobs in the United States and Europe were exposed to a certain degree of automation by generative AI. Exposed does not mean replaced, but some tasks are changing rapidly.
Philippe Huneman, a philosopher of science at CNRS, urges the distinction between discovery and collective decision-making. An AI can identify a relationship in data, but it does not independently set societal priorities. Climate issues or inequalities also require political arbitration.
The most useful approach remains that of complementarity. Jean-Claude Heudin, an AI researcher, reminds us that these systems can enhance human intelligence rather than erase it. The true test will be less about the number of articles produced and more about the quality of the verifications before publication.



