DOE Releases 35 Agent Skills for National-Laboratory Science

The U.S. Department of Energy released 35 technical “skills” that let AI coding agents perform defined national-laboratory tasks, including drafting a job script for Oak Ridge’s Frontier supercomputer and preparing a dataset card.

Twenty-three packages contain executable Python or shell scripts; the other 12 rely on text instructions alone. That split gives research teams automation for some tasks, but no equivalent execution layer for the remainder.

DOE describes the release as “the first building block” of the Genesis Mission’s shared research infrastructure. President Donald Trump has rebranded AI as “SI,” short for “superintelligence,” while DOE uses “Super Intelligence (SI)” following a September executive order requiring federal agencies to replace “artificial intelligence.”

At an Oct. 8 White House summit, 11 companies pledged a combined $2.4 billion in “SI for science tools and compute credits.” OpenAI pledged $200 million, while Google and Anthropic pledged $150 million each. The total averages about $218 million per company, although the source provides no median or spending schedule.

Anthropic, OpenAI and Google signed as Genesis Mission partners and were among 24 organizations that signed memorandums of understanding with DOE. For R&D budget owners, these figures represent commitments rather than awarded contracts or documented disbursements.

What does each skill contain?

Each skill follows the Agent Skills format: a folder built around a SKILL.md file whose metadata tells the agent when to load it. Instructions typically use Markdown, allowing an agent to apply a defined procedure during a research task rather than improvising every step.

The 12 text-only skills include job-scheduling guides for the Frontier, Aurora and Perlmutter supercomputers. DOE says the package complements existing skill libraries focused on coding and software engineering, rather than replacing local operating procedures or validated scientific software.

The catalog covers several operational layers:

• High-performance-computing job schedulers for Frontier, Perlmutter and Aurora
• Data tools, including data cards summarizing what a dataset contains and how it was made
• AI safety analysis workflows and “SI” safety guidance
• Clients for American Science Cloud services, including Globus Compute
• Standalone literature-search and multi-agent-system skills
• Agent evaluation, monitoring, data governance and scientific-workflow guidance

DOE describes these 35 resources as the first technical output from ModCon, formally Transformational AI Models and Data, and the American Science Cloud. The programs build shared AI tools and data for national laboratories under the Trump administration’s initiative to use AI to accelerate scientific research.

Who wrote and reviewed the package?

Attribution files in the repository identify researchers from seven national laboratories. Matt Baughman of Princeton Plasma Physics Laboratory wrote five supercomputer and scheduler skills, while Ian Foster of Argonne wrote the literature-search skill.

Teams led from Lawrence Berkeley and Pacific Northwest national laboratories produced the data and safety skills. The Baseline AI R&D Capabilities team, known as BASE, developed the package within ModCon, with PNNL and Lawrence Berkeley serving as co-leads.

GitHub history records that PNNL’s Nathan Hodas uploaded the first 291 files on Aug. 5 in a commit titled “adding approved genesis skills.” The contribution guide permits AI-assisted work when disclosed and requires a human reviewer on every pull request.

Eight of the repository’s 69 commits list Anthropic’s Claude or GitHub Copilot as co-authors. These controls document review responsibility and disclosure, but the source supplies no error rate, reproducibility result or measure of scientific validity.

Which coding agents can use the skills?

Project documentation identifies compatibility with Claude Code, Codex, Cursor, OpenCode, Gemini CLI and LangChain Deep Agents. DOE warns that individual skills may still depend on client-specific features or external tools, so nominal compatibility does not guarantee identical execution.

For laboratory managers, the relevant test is operational access: the selected agent must receive the required credentials, files, scheduler interfaces and American Science Cloud services. The source does not describe those deployment requirements, supported operating systems or performance under laboratory-scale workloads.

The scheduling guides also do not eliminate the need for institutional controls over compute time, software versions and job priorities. The skills standardize instructions; they do not establish that an agent-generated Frontier script will run efficiently or produce correct results.

How much commercial support is attached?

The $2.4 billion pledge package includes $200 million from OpenAI and $150 million each from Google and Anthropic, earmarked for “SI for science tools and compute credits.” Those commitments could subsidize agent access and compute, but the source gives no allocation by laboratory, project or research area.

Anthropic announced a separate multi-year partnership with DOE in December 2025. The agreement gives national laboratories access to Claude and Anthropic-built tools, including Claude Skills for scientific workflows.

Anthropic’s three-year commitment, announced Oct. 8, makes Claude available to more than 15 agencies, including NASA, the National Institutes of Health and the National Science Foundation. It also provides Claude, Claude Code and API credits for several hundred Genesis Mission projects.

The source does not value those API credits or identify the participating projects. Portfolio managers therefore cannot determine from the announcement alone whether the commitment represents new funding, donated capacity, commercial discounts or a mixture of all three.

What must DOE demonstrate next?

DOE has supplied documented workflows, script-bearing packages and a human-review requirement across a seven-laboratory contribution. What remains unmeasured is execution quality: benchmark success, scheduler compatibility, security boundaries and the proportion of agent-generated outputs accepted without modification.

The next procurement test is whether these packages work across each laboratory’s actual computing stack while preserving reproducibility and access controls. Until those results are published, the 35 skills represent a substantial integration option, not proof of autonomous scientific performance.