What's this about?
The paper treats "agent skills" — reusable instruction packages that customize AI agents — as a proper software artifact class, proposing an ontology and engineering lifecycle (versioning, testing, distribution) for them.
What exactly is an "agent skill"?
A folder of natural-language instructions, metadata, and optional scripts that teaches an agent a workflow — like a recipe card any competent cook can follow. Claude Code's skills are the canonical example.
Why does it matter?
Skills are becoming how agent capability is shared and sold, but they have none of software's hygiene — no semver, no tests, no supply-chain security. Naming the gap is the first step to tooling.