Position: Multi-Agent Scientific Discovery Requires System-Level Safety

As AI agents automate more of the scientific pipeline, safety becomes a property of how agents are organized rather than of any single agent. Capable, well-aligned agents can still compose individually reasonable actions into hazardous outcomes across laboratories, biological design workflows, and security research. We argue that AI-for-science safety therefore requires system-level design for multi-principal, multi-agent scientific ecosystems. We introduce SafeSearch, an abstraction casting scientific discovery as multi-agent search over a space where both value and hazard emerge compositionally, and use it to frame mechanism-design questions for safe scientific coordination.