Los Alamos Hackathon Links AI Agents to Foundation Models

Los Alamos National Laboratory held a winter hackathon this year to connect its agentic AI system with specialized scientific foundation models. About 30 members of the ArtIMis project — the lab's AI for Mission initiative — worked together under the direction of Earl Lawrence, a senior scientist in the Computing and Artificial Intelligence Division Office.

The event aimed to link the Universal Research and Scientific Agent, or URSA, to other AI products developed through ArtIMis. URSA is an agentic architecture designed to help scientists answer research questions or handle repetitive tasks without constant human input. The hackathon participants integrated scientific foundation models — pre-trained AI systems unleashed on large Lab datasets — into URSA's knowledge repository.

According to the official LANL news release, the integration makes the framework more capable of answering the wide array of questions Lab scientists might pose. These foundation models can now identify patterns to accomplish complex tasks they were not initially trained for.

URSA can summarize literature, write code, run scientific simulations, generate plots, hypothesize new simulations, or draft papers as it works toward an answer. Nathan DeBardeleben, a senior research scientist and co-principal investigator for ArtIMis, likens using URSA to a mentor giving a mentee a problem to solve. From the initial text query, a user might have to provide hints to guide URSA toward a desired outcome.

"It works toward a solution and then finally stops when it thinks it has met your criteria," DeBardeleben says. "A student is a good approximate — sometimes you'll think you defined the problem incorrectly to the student, other times the student stops before you wished they had, other times it finds something interesting and goes off on a bit of a tangent. You can pick up the pieces whenever you want with URSA at these checkpoints and steer it back on course."

That checkpoint control matters when you're actually using the system. Scientists don't just type a query and walk away — they monitor progress, intervene when the agent drifts, and validate outputs before trusting them for mission work. The physical reality involves clicking through checkpoints, reviewing generated code, and verifying simulation parameters. It's not magic; it's a tool that requires oversight.

With nearly 100 Lab employees contributing to the ArtIMis project, collaboration across AI-focused capability-building teams and scientific-application teams is no easy task. Emily Casleton, a scientist and team leader of the ArtIMis testing and evaluation team, identified the hackathon as an opportunity to get members of different teams in one location at one time to build bonds across the groups.

"A big thing that came out of it was just seeing who's doing the model building, who's doing the testing and evaluation," Casleton says. "Now, if anyone has a question about something, I know who to go to."

Casleton's team was charged with ensuring the foundational models were connected to URSA and running properly. As the team lead, she was the hackathon's "final boss." Lawrence noted that Casleton said before everyone leaves, she has to be able to run their model. "I don't know if we quite achieved that, but regardless, everyone made such tremendous progress that I really think the productivity was quite high."

The hackathon also serves the broader Genesis Mission, a Department of Energy-wide AI effort. Lawrence serves as the National Nuclear Security Administration's models pillar lead for Genesis. The process of making foundational models accessible to URSA serves as a proxy for what is expected to happen on a national scale.

"If we want someone at another lab or a large DOE-wide ecosystem to be able to access the things we developed, it's the exact same process," Lawrence says. "So, in my mind, that was always an ulterior motive. If we solve the problem and make sure our test and evaluation team can run stuff, we also contribute to the national initiative."

Independent coverage from HPCwire confirms the timeline and scope of the hackathon, noting it occurred over the winter of 2026. The outlet's reporting aligns with LANL's official account of the event's goals and outcomes.

At a more insular level, the hackathon represents a major step in providing more scientists with a robust AI tool for use across the Lab's mission spaces. Lawrence hopes to host a hackathon or similar event every quarter so the ArtIMis project can maintain a cohesive thrust and direction.

"There's a lot of hype and a lot of excitement around AI, particularly with Genesis, but it was nice to see the nuts and bolts of this all come together, and how scientists and AI experts can work side by side to solve real problems," Casleton says. "It's not just fluff, and that's really exciting for me."

Whether quarterly hackathons can sustain momentum across a 100-person project remains to be seen. The real test comes when these integrated systems handle actual mission-critical work without constant human intervention. For now, the checkpoint model means scientists still need to be in the loop — which is probably the right call given the stakes involved.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn