23% of Agent Skills Already Bundle Executable Code, Not Just Prompts
Notes on Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts (arXiv:2607.18970) — Haodi Fan, Zucong Lan · July 2026
Note published · written by SkillFed’s research pipeline from the paper above · how these notes are made
AI-assisted notes · reviewed by SkillFed Skill evolutionSkillware is Fan and Lan's name for what an agent skill actually is once you stop treating it as a prompt: a three-layer object. The Skill Artifact is just the natural-language task spec. Wrapped around it sits the Skillware Unit — an independent software identity with its own addressable name, version, compatibility record, and provenance trail. The Agent Host is whatever runtime discovers, activates, and interprets the thing. Membership in the category isn't just declared; it's tested — three necessary conditions (behavioral primacy, independent software identity, compatible host execution) checked against 15 boundary cases, 12 that should count and 3 deliberately chosen not to. The evidence behind it is a frozen corpus, SkillMD-138K: 138,133 content-deduplicated SKILL.md files pulled from 20,556 repositories, plus 13 fixed-revision case studies of real implementations.
The convention is already locked in without anyone mandating it: 98.73% of files carry structured frontmatter, reading like package manifests years before a formal spec showed up to describe them. More telling is what's inside — 23.2% of files, 32,069 of them, point explicitly into references/, scripts/, or assets/ directories, so close to a quarter of skills in the wild have already outgrown plain prompt text for resource- or tool-backed bundles. Independent studies cited in the evidence sharpen the risk side: one traced 3,709 reuse links across 18,463 registry Skills and 23,199 personal-use Skills and found copies get modified after reuse rather than staying fixed; another sampled 238 Skills and found quality smells were widespread. Self-evolution, meanwhile, is defined as the framework's most demanding dimension — a governed mechanism that must propose, evaluate, and adopt changes while preserving identity — and Lifecycle Continuity (whether a unit's identity survives update, rollback, and removal) gets tracked precisely because it can't be assumed yet.
Key numbers
| SKILL.md files analyzed | 138,133 (20,556 repos) |
| Files with structured frontmatter | 98.73% |
| Files bundling scripts/refs/assets | 23.2% (32,069 files) |
| Skills sampled with widespread quality smells | 238 |
Skills related to this research
Related notes
- Zero of 8 production agent systems verify skill-in-use →
- Code review, testing, and security auditing claim 35% of task assignments; requirements analysis gets 2% →
- Agent-skill catalogs already top 700,000 entries — curation hasn't caught up →
- Flat retrieval breaks once a skill library hits the tens-to-hundreds range →
- 215 Skills, 165 Contributors, No Fidelity Test →
- 26.1% of Community Skills Ship With a Vulnerability →
- Remove the curriculum and Voyager's item discovery collapses 93% →
- Retrieving from a 210-skill library mines diamonds 2.5x faster than writing code on the fly →
- Retrieving the right skill lifts coding Pass@1 78%, reaching 84% of oracle-level performance →
References
- Fan, H. & Lan, Z. (2026). Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts. arXiv:2607.18970.