Speaker
Description
Operating large-scale scientific facilities requires coordinating diverse subsystems, translating operator intent into precise hardware actions, and maintaining strict safety oversight. Language-model agents offer a natural interface for these tasks, but most existing approaches are not yet reliable or safe enough for production use. We introduce Osprey, a framework that wraps a coding agent in a control-room operator interface, a tool surface that reaches hardware through pluggable connectors for the control system used in our community, and a first-class component for natural-language search of facility electronic logbooks. The agent itself is treated as a replaceable component: operator interface, safety policy, tool servers, and connectors stay under facility control, while the agent backend can be swapped as the AI ecosystem evolves. A declarative build-profile mechanism lets each facility maintain its own configuration without forking the shared framework, keeping deployments reproducible across updates. Osprey has been deployed at several DOE accelerator facilities through the MOAT seed effort within the Genesis~Mission. This paper presents the current framework architecture and reports on the substantial evolution Osprey has undergone over the past year.
| Paper status | Resubmitted proceeding files received and assigned to an editor. Accepted by Submitter. |
|---|