Rescale's Engineering Agents Go to Work on a Real Nuclear Reactor

Kairos Power projects that agent-driven simulation workflows could triple end-to-end analysis efficiency on its Hermes 2 reactor program. Rescale, the cloud simulation company, put that number on the table on November 10, 2026, when it introduced a framework for customizing engineering AI agents and named its first high-stakes customer pair: Kairos Power and Oak Ridge National Laboratory's Manufacturing Demonstration Facility.

The two organizations are using Rescale's platform for Hermes 2, a 50-megawatt fluoride-salt-cooled reactor. The collaboration is testing advanced manufacturing and construction methods. Rescale's AI developer engineers supported the work.

What are the agents actually doing?

The teams deployed agents for thermal, flow and shielding simulations, plus the reporting that goes with them, on reactor-cavity and bioshield-plug designs. An engineer approved every agent action. That detail is the story's center of gravity. This is not autonomous AI loose inside a nuclear program; it's supervised machine work with a human signature on each step.

ORNL's contribution is institutional knowledge made portable. The lab is encoding its simulation and additive-manufacturing expertise into agents, so engineers beyond the Manufacturing Demonstration Facility can adopt those methods.

The efficiency claim is prospective, not delivered. Kairos Power projects that agentic workflows built from past simulations could lift end-to-end simulation analysis efficiency by up to 3x. The condition attached matters as much as the multiplier: the gains must come "while preserving the traceability needed for regulatory compliance." In nuclear engineering, an unauditable result is a worthless result, no matter how fast it arrived.

Why does this matter beyond one reactor?

Nuclear licensing runs on documented evidence. If reusable agent workflows can cut analysis effort while keeping an auditable engineering record, the same pattern could spread across regulated industries — aerospace, medical devices, energy — where the bottleneck is not computing power but the time engineers spend producing defensible documentation. The Hermes 2 collaboration is effectively a live test of that proposition.

How do Agentic Projects work?

Rescale's Fall 2026 Showcase made Agentic Projects the control layer for Edison, the company's AI assistant and agent orchestrator. Teams set objectives, tools, skills and guardrails for a project, then add agent nodes for individual steps:

• validating inputs
• analyzing results
• preparing evidence

Projects can be saved and shared across programs. When Edison proposes a learning, it enters project memory only after an engineer accepts it. Nothing the system learns becomes part of the record without a human clicking yes.

Engineering Intelligence, a separate layer, links requirements, design, simulation, test and process records in knowledge graphs. That gives agents the relationships and configurations they need for complex work. A work graph runs alongside it, mapping people, data and tasks across a workflow, so engineers can trace how a piece of work was performed and exactly where agents participated.

What was already in place before the showcase?

The November 10 event was not the first appearance of every component. Rescale's Agentic Projects release note is dated October 26, 2026, and describes:

• validated baseline jobs
• linked standards and files
• reusable skills
• citations when Edison answers from a linked document

Projects begin private and can be shared with a workspace. Enabling them requires Agentic Engineering to be switched on.

Rescale Desktop followed in a November 5, 2026 release note. The desktop app runs on macOS for Apple silicon, Windows x64 and Linux x64. It moves files between local and cloud environments, tracks jobs, and provides Edison chat without a browser.

The desktop client carries its own privacy boundaries. Edison can search only user-approved folders. It can see file names, sizes and dates — but not file contents. It asks before every transfer. Chat requires an Advanced Modeling & Simulation or Agentic Engineering subscription, and it is not yet available on Rescale ITAR, the deployment tier for export-controlled work.

How do outside AI tools plug in?

Rescale also expanded its Model Context Protocol connectors, which let an outside AI client issue natural-language requests to the platform. The company says the connectors expose its specialized agents and more than 1,350 pre-installed engineering applications inside Claude and other MCP-compatible tools. Rescale MCP is listed in the Claude and Codex directories.

The company drew a firm line around that route. The MCP documentation limits it to interactive use by a person working through an agent. Each connection binds to one workspace and inherits the signed-in user's existing permissions, with actions attributable to that user. For production integrations, Rescale directs customers to its REST API or command-line interface.

That is a deliberately narrow door. The MCP path is for engineers conversing with their tools, not for headless pipelines acting on their own.

What changed in AI physics?

The showcase also added AI physics features aimed at surrogate models — AI approximations trained from simulation data rather than first-principles computation. A dataset-creation agent turns existing simulation data into training-ready datasets. Guided pre-processing identifies outliers and coverage gaps and recommends data splits. Auto Tune adjusts model configurations during training.

Design Explorer provides a guided local interface for evaluating thousands of design options against constraints set by simulation engineers. It can also link multiple surrogate models together into multiphysics optimizations.

Uncertainty quantification attaches confidence and out-of-distribution signals to each prediction. It flags results that need a full-fidelity simulation check and identifies cases where the model itself needs expansion. In a domain where a wrong number can invalidate a safety case, knowing when not to trust the approximation is a feature, not a caveat.

Is a human still in charge everywhere?

Yes — and the release applies that boundary consistently across three layers. Agent-suggested learnings require engineer approval before entering project memory. Hermes 2 agents act only with human sign-off on every action. Surrogate predictions that carry uncertainty get flagged for full-fidelity checks rather than accepted as-is.

The Hermes 2 collaboration is the live experiment sitting on top of all this machinery. It tests whether reusable agent workflows can reduce analysis effort while preserving an auditable engineering record — and whether the 3x efficiency projection holds up against the compliance demands of nuclear regulation.