Beyond Centralized Labs: Federating the Co-Scientist

AI co-scientists have demonstrated that agentic AI systems can make meaningful contributions to scientific discovery; however, existing designs are largely centralized and tied to a single site or computing environment. This paper introduces a federated co-scientist architecture that extends the concept of co-scientist to a distributed setting. Built on the Academy framework, we present a generic design that can be instantiated as a first generation of federated co-scientists. Instead of a single agent or a cluster of co-located agents, we deploy a network of federated agents that can be placed near instruments for low-latency control, in HPC facilities for large-scale computation, adjacent to large data stores for high-throughput access, and in the cloud for global coordination. A federated setting enables these agents to exchange hypotheses, models, and summaries without centralizing raw data or even control, supporting scalable, policy-aware collaboration across institutions while preserving sovereignty. Crucially, the federated paradigm supports the extension of co-scientists into specialized teams that learn within communities and domains and collectively advance scientific discovery.