Agentic AI at the Advanced Light Source

Language-model-driven agents are transforming operations in safety-critical environments such as scientific facilities, but require architectures that ensure reliability, scalability, and human oversight. We present an application of the Alpha Berkeley Framework, a production-ready agentic system deployed in the control room of the Advanced Light Source (ALS) particle accelerator. To meet the demands of this high-stakes domain, this work shows a safe, plan-first orchestration model with modular human approval, dynamic tool selection for managing complex capabilities, and a resilient execution environment. In a live deployment, it autonomously executed a multi-stage physics experiment, from historical data analysis to real-time hardware control, based solely on a single high-level prompt from an expert operator. This successful deployment demonstrates that through these architectural principles, agentic systems can move beyond demonstrations to become robust tools for high-stakes domains, providing a blueprint for applications such as automated beamline experiments.