AI bots wrote and reviewed all papers at this conference (www.nature.com)

Agents4Science 2025 will be the first scientific conference in which every submitted paper and every review was generated by AI agents. The online event (22 October) will feature presentations delivered either by the agents themselves or by the humans who ran them, with humans allowed only advisory roles while AIs serve as the primary contributors. Organizers received submissions from over 300 agents and accepted 48 papers—mostly computational studies spanning topics from psychoanalysis to mathematics—with the review process also carried out by AI reviewers. To enable analysis, submitters had to document the nature and extent of human interaction at every step. The event is a deliberate sandbox to study a fast-emerging research paradigm: coordinated groups of models (agents) functioning as “scientists” rather than standalone tools. That shift raises technical and policy questions central to AI/ML: how to evaluate agent-driven research, how to quantify and mitigate false-positive or hallucinated discoveries, and how human oversight level affects quality. The conference will produce empirical data on agent capabilities and failure modes that could inform publishing rules, reviewer workflows, and standards for agent transparency. Organizers also hope such venues might reduce reviewer load at traditional conferences while clarifying the ethical and methodological boundaries of machine-authored science.