NewsBB-0098

Paper: an LLM agent writes ARTIQ control code for a trapped-ion experiment behind hardware safety tokens

Researchers put an LLM agent in charge of a trapped-ion experiment's ARTIQ control code through an MCP server. Every tool call needs a token bound to its contents, issued by a human or after a simulated run stays within device limits. The paper reports that the agent built a calibration stack and, with operator guidance, closed a magnetic-field stabilization loop.

Open original source arxiv.org · Duanyang Wang, Lu Qi, Yuanheng Xie, Norbert M. Linke, Kenneth R. Brown

Claimed
Why this evidence label?

Paper reports results on the physical trapped-ion apparatus; no code release linked.

How we check entries
Preview image for Paper: an LLM agent writes ARTIQ control code for a trapped-ion experiment behind hardware safety tokens
The source's own preview image, from arxiv.org.

Briefing

Each script first passes an abstract-syntax-tree check against a denylist that includes os.system and subprocess, then runs in dax.sim, a simulator for ARTIQ hardware. Every DDS setting, DAC voltage and TTL pulse in the trace is checked against per-device bounds, and an unmapped device goes to a human. The agent was built on Claude Opus 4.8 and kept its goal and progress in Markdown files between sessions. The stabilization loop suppressed 60 Hz power-line noise.

The authors list the gaps: deliberately rewritten code can slip past the syntax check, and the simulator does not model every driver feature. The operator stepped in each time the agent kept refining a method that had run its course, such as a Gaussian decay model. The portability test ran on a 171Yb+ experiment at the University of Maryland.

Tags

Interface

Report a problem with this entry