Lawrence Livermore National Laboratory (LLNL) has deployed what it bills as the Multi‑Agent Design Assistant (MADA) to speed inertial confinement fusion (ICF) target design. The system couples large language models with LLNL’s 3D multiphysics code, MARBL. The goal? Automating the generation of simulation decks.
Researchers run the agent framework on the El Capitan supercomputer, among the world’s fastest at 2.79 exaFLOPs peak, and its smaller sibling Tuolumne. The approach uses an “Inverse Design Agent” to translate hand‑drawn capsule diagrams into thousands of simulations and a “Job Management Agent” to schedule runs across the HPC systems.
The project is the result of years of development, with origins dating back to 2019. LLNL physicist Jon Belof, the principal investigator, said the team initially explored combining AI with shockwave physics, what many considered “a strange idea” at the time, Belof said in a press release. In a recent demonstration, an open-source LLM fine-tuned on MARBL documentation took a hand-drawn capsule diagram and natural language request from a human designer, then produced a complete simulation deck and ran thousands of simulations exploring variations in ICF capsule geometry.
Belof argues that AI agents could compress design cycles drastically, allowing researchers to examine not just 3–4 distinct ICF design concepts but hundreds or thousands in parallel.
Rather than the human running ensembles of simulations, they will be able to run ensembles of ideas. —Belof
After LLNL’s December 2022 ignition milestone, the lab is now hunting for a robust high-gain platform for national security applications.
Two agents working in concert
The MADA framework incorporates two key AI agents working in tandem. The Inverse Design Agent tackles design generation, while the Job Management Agent drives execution of large-scale simulation workflows, interacting with the Flux scheduler and workflow management tools like Merlin. “The Job Management Agent brings AI and HPC together to coordinate agents that handle resource management and workflow optimization at massive scales,” said JMA Team Lead Giselle Fernandez.
The approach has already produced promising results. The team tapped the Tuolumne supercomputer to run ensembles of tens of thousands of ICF simulations in single studies. These outputs train a machine learning model called PROFESSOR, which generates instant feedback for designers. “Once trained, the PROFESSOR model generates implosion time histories, radius as a function of time, that change instantaneously when the human designer changes the input geometry,” Belof said.
Funding for the project came from the National Nuclear Security Administration’s Advanced Simulation & Computing program. The MADA team includes LLNL researchers Charles Jekel, Rob Rieben, Will Schill, Meir Shachar, and Dane Sterbentz, along with collaborators Nathan Brown from Sandia National Laboratories and Ismael Djibrilla Boureima from Los Alamos National Laboratory. Belof notes the system represents something unprecedented: “We are putting AI in the driver’s seat of a supercomputer, which is something that has never been done before.”
Federal agencies test the agentic waters – with mixed results
LLNL’s push comes as other government agencies experiment with similar “agentic” AI systems, but not all reviews are glowing. In June 2025 the FDA rolled out “Elsa,” an AI assistant reportedly trained on Anthropic’s Claude. The tool drew mixed reactions. Some staff told Regulatory Focus that Elsa helps them parse test reports and cuts review times, while others complained that its knowledge stops in April 2024 and that it has given outdated answers about agency leadership. Elsa carries a disclaimer warning users to verify its responses because “Elsa can make mistakes.”
Meanwhile, NASA’s Goddard Space Flight Center is working on “text to structures” and “text to spaceship” projects that let scientists describe desired spacecraft designs and have AI propose lightweight, efficient structures. Omar Hatamleh, Goddard’s chief AI officer, said the agency is experimenting with agentic workflows to identify bottlenecks in procurement and finance functions and remove rote tasks from human workers.
A January 2025 Nextgov/FCW column predicted that AI agents will automate back‑office tasks and optimize workflows in government, freeing up human staff for higher‑value work. The piece argued that AI‑powered chatbots will provide constituents with 24/7 access to information and highlighted the U.S. Patent and Trademark Office’s AI‑assisted search system as an example of how agencies can leverage AI to improve efficiency and accuracy.
Industry hype meets organizational reality
Consultancies are selling a grand narrative. Capgemini’s July 2025 report “Rise of agentic AI” pegs the economic opportunity at up to $450 billion in revenue growth and cost savings by 2028. But the same survey found that only 2% of organizations have deployed AI agents at scale and that trust in fully autonomous agents fell from 43% to 27% in one year. The report lists prerequisites such as redesigning processes, reimagining business models and balancing autonomy with human oversight.
McKinsey’s January 2025 report “Superagency” argues that agentic AI now can converse with customers and autonomously handle follow‑up steps like processing payments and checking for fraud. The consultancy notes that software vendors are embedding agentic capabilities into platforms to create “digital workforces”, yet only 1% of companies consider themselves mature in deploying AI.
