What's this about?
Scientific datasets are increasingly consumed by AI agents, but existing dataset representations offer little support for autonomous discovery, interpretation and invocation. This packages scientific data services as agent-ready skills.
What exactly is missing from a dataset for an agent?
A dataset is usually published for a human: a download link, a paper describing it, a README with column meanings. A human reads all that, understands the caveats, and writes code. An agent needs the same information in a form it can act on without a human intermediary — what this contains, what questions it can answer, what its units and limitations are, and how to query it. That's the difference between publishing data and publishing a service. And "skill" is the apt packaging, given this week's finding that skills work by encoding procedure — how to correctly use this dataset is exactly procedural knowledge.
Why is this worth noticing?
- It's infrastructure for the AI-scientist agenda that most of the flashier papers assume already exists.
- It generalizes past science: the same gap exists for every internal company dataset an agent is expected to use.