Large-language-model (LLM) agents are increasingly capable of writing and running experimental control code, opening the door to autonomous laboratories. But giving an AI direct access to hardware also creates a fundamental safety challenge: how can we allow useful autonomy without giving the agent unrestricted control? We present a safety-aware control system that places an LLM agent in the loop of a trapped-ion experiment while enforcing strict authorization for every hardware operation. Proposed actions are either automatically verified in simulation against predefined device limits or explicitly approved by a human operator. Within these constraints, the agent can design, execute, and refine its own experiments. We demonstrate the system on trapped-ion platforms, where the agent autonomously develops calibration procedures and assists with magnetic-field stabilization. We also test the authorization framework against adversarial attempts to bypass its safeguards. Our results highlight both the promise of autonomous experimental control and an important limitation of current agents: knowing when to rethink an experimental problem.

Pizza and drinks will be served after the seminar in ATL 2117.