Three specialized AI agents are now running fusion design experiments with almost no human in the loop.
In a paper posted to arXiv, researchers describe MADA (Multi-Agent Design Assistant), a framework that splits a scientific design workflow across three large language model agents: one manages simulation jobs on high-performance computing systems, one generates the meshes those simulations need, and one proposes new design tweaks based on what the results show. The team tested MADA on Richtmyer-Meshkov instability suppression, a problem in inertial confinement fusion where turbulence at material interfaces can wreck a fusion capsule's implosion. They ran it two ways: against real hydrodynamics simulations on HPC systems, and against a pre-trained machine learning surrogate for faster iteration. In both setups, MADA iterated on its own designs and moved them toward better instability suppression without a person manually reconfiguring each run.
The interesting part isn't the fusion result itself, it's the plumbing. Most "AI for science" pitches show a model spitting out a prediction; this one wires reasoning, simulation, and specialized tools into an actual iterative workflow, which is the tedious part that normally eats a researcher's week.
It's a proof of concept, not a fusion breakthrough, and the paper is explicit that RMI suppression was just the test case for a pattern they want reused elsewhere.