Agents4Science: A Conference Where AI Acts as Lead Scientist

Agents4Science will debut in October as a one-day online academic event that treats AI systems not just as tools, but as primary researchers. The conference invites submissions where AI is the first author, the main researcher, and the principal driver of experiments and papers.

Background and ambitions

The event was organized by Stanford computer scientist James Zou, who studies human-AI collaboration in research. Advances in large language models and reasoning-enabled AI have moved the idea of autonomous AI scientists from speculation toward practical demonstration. Tools like AlphaFold already assist researchers in domains such as protein modeling; Agents4Science aims to test how far AI can go when it takes the lead in proposing hypotheses, running simulations, and designing experiments.

The Virtual Lab concept

Zou and collaborators developed a concept called the Virtual Lab: a team of AI agents that simulate the roles found in a university lab. Each agent can be trained with different expertise and given access to programs and resources, allowing them to coordinate on a research agenda and propose experiments for human follow-up. Zou teamed up with John E. Pak from the Chan Zuckerberg Biohub to build and test such a system.

A rapid research run: nanobodies and covid

For their first project, the Virtual Lab focused on designing therapies for new covid-19 strains. Zou trained five AI agents with roles like immunologist, computational biologist, and principal investigator. Once configured, the agents quickly generated candidate therapies — Pak estimated the models produced designs within a day or less. The agents selected anti-covid nanobodies, smaller cousins of antibodies, partly because those molecules fit the computational constraints the team provided. Many of the designed nanobodies bound to the original covid-19 variant, and the paper emphasized that the Virtual Lab itself is the primary contribution as a tool for automated discovery.

Policy, credit, and the motivation for Agents4Science

Conventional journals and conferences often forbid listing AI systems as authors or crediting them as primary contributors. Nature, for example, has cited concerns about accountability, copyright, and inaccuracies. Zou sees those restrictions as incentivizing researchers to hide or minimize AI's role. Agents4Science flips that norm by requiring the primary author on submissions to be an AI and using other bots to attempt peer review. Human experts, including a Nobel laureate in economics, will still review top submissions.

Supporters and critics

Some scientists and policymakers are intrigued by the potential of AI scientists. The US government's AI Action Plan even mentions investment in automated cloud-enabled labs. Proponents argue that AI agents can work continuously and explore spaces humans cannot, possibly uncovering discoveries humans might miss.

Skeptics raise strong concerns. Critics question whether AI can achieve the leaps of insight that characterize major scientific breakthroughs and worry about hallucinations, errors, and the risk that automation could limit training opportunities for junior researchers. Scholars like Lisa Messeri and Molly Crockett argue for involving philosophers, epistemologists, and anthropologists to clarify what counts as knowledge and to design thoughtful experiments before trusting AI to perform and evaluate science.

An experiment in public

Zou frames Agents4Science as an experiment intended to produce systematic data on what AI-driven research can and cannot do. He hopes the conference will surface both promising results and instructive mistakes. In October, autonomous agents will present their work via text-to-speech and AI-authored papers will be evaluated by bot reviewers and human experts, offering the research community an opportunity to see the current capabilities and limits of AI scientists in practice.