SkillWiki: Living Infrastructure for Agent Skills
- SkillWiki is a living knowledge infrastructure that transforms raw artifacts into executable agent skills, integrating provenance and lifecycle governance.
- It separates a knowledge layer from a skill layer, using a provenance graph to link knowledge artifacts with evolving capabilities across atomic, functional, and strategic levels.
- The platform employs a Git-style governance workflow and execution-driven evolution to continuously monitor, refine, and maintain agent skills.
SkillWiki is a “living knowledge infrastructure” for agent skills: a platform that does for skills what Wikipedia does for textual knowledge and GitHub does for software. It is designed for ecosystems of LLM-based agents that must accumulate, share, and continuously refine skills over time, rather than for a single agent’s private memory. In SkillWiki, a skill is explicitly treated as a lifecycle-aware capability asset bundling identity and classification, lifecycle state, interface specification, implementation, evaluation contract, provenance, graph relations, and runtime metrics (Huang et al., 15 Jun 2026).
1. Conceptual model
SkillWiki separates a knowledge layer from a skill layer. The knowledge layer is a long-term repository of raw materials—trajectories, documents, API specs, scripts, historical skills, and agent execution records—while the skill layer contains executable skill assets derived from those materials and linked back to them as evidence (Huang et al., 15 Jun 2026). This separation is foundational: knowledge artifacts are preserved as source evidence, whereas skills are governed, versioned, and monitored as operational capabilities.
The platform centers on a Skill Provenance Graph. Conceptually, it is a heterogeneous directed graph whose nodes include knowledge nodes, skill nodes, execution nodes, validation or governance nodes, and lifecycle events, and whose edges include relations such as derived_from, refines, depends_on, executed_in, and validated_by (Huang et al., 15 Jun 2026). The graph is not only a storage structure; it is the mechanism that makes provenance-aware exploration, governance, impact analysis, and version lineage possible.
SkillWiki also classifies skills by taxonomy level: atomic, functional, and strategic (Huang et al., 15 Jun 2026). This aligns it with adjacent work in which “skills are represented as nodes in a directed graph, with typed edges encoding prerequisite, enhancement, and co-occurrence relations” (Li et al., 12 May 2026), and with automated skill knowledge bases that partition skills into planning, functional, and atomic levels (Wang et al., 6 Apr 2026). This suggests that SkillWiki is best understood as an infrastructure layer that can host multiple skill formalisms rather than a single fixed representation.
2. Knowledge-grounded skill production
SkillWiki’s production workflow transforms heterogeneous artifacts into governed skills. The supported knowledge sources are trajectories, documents, API specifications, scripts, historical skills, and execution records (Huang et al., 15 Jun 2026). The pipeline parses and normalizes each artifact, uses an LLM to identify reusable actions and workflows, generates candidate skills with interface and implementation drafts, registers provenance links, and places the resulting artifacts into the lifecycle (Huang et al., 15 Jun 2026).
The evaluation corpus includes 125 artifacts, 25 per category (Huang et al., 15 Jun 2026). In that study, 99 out of 125 artifacts yielded at least one governed skill candidate (Huang et al., 15 Jun 2026). The per-source breakdown is explicit: trajectory artifacts produced 24 candidate and 24 governed skills; document artifacts 17 and 17; API specifications 24 and 24; scripts 18 and 18; historical skills 16 and 16 (Huang et al., 15 Jun 2026). The production success is therefore high for trajectories and API specifications, while documents and historical skills are more heterogeneous.
The generated skill is not just a free-form prompt. A candidate skill contains a draft interface, a draft implementation, an initial evaluation contract, and links to source artifacts (Huang et al., 15 Jun 2026). Related systems expose why this structure matters. SkillCom decomposes LLM-based semantic communication into “semantic abstraction skill, channel-adaptive transmission skill, receiver-side repair skill, and task execution skill,” with typed semantic-unit interfaces (Fu et al., 4 May 2026). SkCC, in turn, treats SKILL.md as a compilable source artifact and introduces “SkIR - a strongly-typed intermediate representation” plus compile-time analysis and platform-specific emission (Ouyang et al., 5 May 2026). Together, these results indicate that skill production increasingly requires typed interfaces, explicit dependencies, and provenance rather than unstructured prompt text alone.
3. Provenance-aware repository and exploration
SkillWiki’s repository is both a catalog and a graph-backed exploration system. Skills are indexed by taxonomy, lifecycle state, interface, implementation, provenance, dependencies, and health (Huang et al., 15 Jun 2026). Search can therefore operate on multiple axes: by domain, by input/output signature, by source artifact, by dependency structure, or by execution health. This is substantially richer than flat semantic-similarity retrieval.
The provenance surface is central. Skill detail views expose the source artifacts from which a skill was derived, the skills that depend on it, the versions that refine or supersede it, and the executions and reflections associated with it (Huang et al., 15 Jun 2026). This makes a skill auditable as an evolving asset rather than a static instruction block.
Earlier semantic-wiki systems help clarify the lineage of this design. SHAWN made “every page into a concept in a semantic network, and every ‘field: value’ pair into an RDF-like triple” (Aumueller, 2010). NNexus provided “automatic linking for previously defined concepts” through concept indexing, concept discovery, and link annotation (Ginev et al., 2014). SkillWiki inherits the same concern with interlinking knowledge, but the units are now executable skills, knowledge artifacts, execution records, and governance events rather than encyclopedia entries alone. This suggests that SkillWiki extends semantic-wiki ideas from descriptive knowledge into provenance-aware operational knowledge.
4. Governance, lifecycle, and versioning
SkillWiki defines a lifecycle with eight states (S0–S7) and manages them through a Git-style governance workflow (Huang et al., 15 Jun 2026). Changes are represented as proposals rather than in-place edits; each proposal includes structured diffs, impact analysis on dependencies, supporting evidence, and automated checks (Huang et al., 15 Jun 2026). Governance can run autonomously by default, but human users can inspect proposals, override or approve them, and force transitions in lifecycle states (Huang et al., 15 Jun 2026).
| State | Name | Meaning |
|---|---|---|
| S0 | Raw Experience | Ingested knowledge or execution trace not yet structured into a skill |
| S1 | Candidate | Candidate skill generated from knowledge artifacts |
| S2 | Draft | Under active refinement and validation |
| S3 | Verified | Passed validation tests and governance review |
| S4 | Released | Official, recommended version used by agents |
| S5 | Degraded | Monitoring detects issues |
| S6 | Deprecated | Superseded or judged unsafe or obsolete |
| S7 | Archived | Fully retired but preserved for traceability |
A typical trajectory begins with ingestion and candidate generation, moves through draft and verification, reaches release, and may later pass into degraded, deprecated, and archived states as runtime evidence accumulates (Huang et al., 15 Jun 2026). A full-chain case study shows a skill derived from API documentation progressing through all states S0–S7 (Huang et al., 15 Jun 2026).
This governance design addresses long-standing wiki concerns. Earlier work on wikis identified problems including “identification of the identities of information and its creator,” “accuracy of information,” “vandalism of quality of information,” and “weak control over the contents,” and proposed an annotation model with “systematic control of contributors and contributions” and “keeping of records of what existed and what was done to initial documents” [0702106]. Another account described wikis as “anarchistic,” “collaborative,” “connected,” “organic,” “self-healing,” and “based on trust,” while stressing tensions between openness and organizational control (0802.0745). SkillWiki does not adopt anarchic editing as its core principle; it replaces that model with provenance, lifecycle control, and proposal-based governance.
The contrast with enterprise skills tools is also instructive. In a software consulting company, a “skills manager” supported resource allocation, expertise finding, identifying new project opportunities, and skills upgrading, and its usefulness depended on visible incentives, structured competence fields, and regular updating (Dingsøyr et al., 2019). SkillWiki extends that logic from human competence directories to executable agent capabilities, but preserves the same emphasis on versioning, discoverability, and organizational use.
5. Execution-driven evolution
SkillWiki’s “living” aspect is the execution loop. When agents use a released skill, the system logs execution statistics, failure evidence, reflection memories, and lifecycle signals (Huang et al., 15 Jun 2026). The paper defines success rate and failure rate as
and
for a skill (Huang et al., 15 Jun 2026). These signals feed monitoring, degradation detection, and maintenance proposal generation.
From execution evidence, SkillWiki can propose repair, decomposition, replacement, or retirement (Huang et al., 15 Jun 2026). Reflection memories and recurring failure patterns become new knowledge artifacts, so knowledge updates, skills, and execution experience co-evolve in a single infrastructure (Huang et al., 15 Jun 2026). This closes the loop that many adjacent systems handle only partially.
Several contemporary systems illuminate this evolution layer. SkillAxe “decomposes skill quality into four interpretable dimensions” and, on SpreadsheetBench, “raises pass rate from 16.0% to 52.0% using only 22 skills” (Gautam et al., 9 Jun 2026). SkillX constructs a “plug-and-play skill knowledge base” through “Multi-Level Skills Design,” “Iterative Skills Refinement,” and “Exploratory Skills Expansion” (Wang et al., 6 Apr 2026). SkillTTA retrieves a small set of trajectories and synthesizes “a temporary, task-specific textual skill,” improving SpreadsheetBench Pass@1 from 0.397 to 0.505 and BigCodeBench Pass@1 from 0.517 to 0.651 (Wang et al., 16 May 2026). SkillGraph continuously updates a graph-structured library from trajectories and reinforcement-learning feedback (Li et al., 12 May 2026). SkillWiki can be read as the infrastructure needed to host, govern, and compare such evolving skill assets over time.
6. Demonstration, positioning, and open problems
SkillWiki includes a web-based UI and a CLI, and the demonstration exposes four workflows: knowledge ingestion, skill exploration, skill governance, and execution-driven evolution (Huang et al., 15 Jun 2026). The system is positioned as a missing infrastructure layer: “while knowledge is managed through Wikipedia and software through GitHub, agent skills still lack an infrastructure for large-scale production, governance, and evolution” (Huang et al., 15 Jun 2026). In that sense, SkillWiki is neither only a benchmark nor only a memory module; it is an attempt to operationalize skills as shared infrastructure.
The broader ecosystem reinforces this positioning. In telecommunications operations, SKILLS shows that a portable [SKILL.md](https://www.emergentmind.com/topics/skill-md-8e654683-8cbe-45f8-b1a6-aff3b05133e7) document encoding workflow logic, API patterns, and business rules delivers consistent “skill lift” across models, with MiniMax M2.5 reaching 81.1% with-skill and Nemotron 120B and Seed 2.0 Lite each gaining +18.9 percentage points (Brett, 16 Mar 2026). SkCC shows that SKILL.md has become a de facto artifact format, and that a compilation framework can reduce adaptation complexity from to while adding security analysis (Ouyang et al., 5 May 2026). These results suggest that SkillWiki’s repository, governance, and provenance mechanisms are timely precisely because skills are becoming standardized, portable artifacts rather than ad hoc prompts.
The current system also has explicit limitations. The evaluation is on 125 artifacts and a modest skill set; behavior on repositories with “tens of thousands of skills” remains unexplored (Huang et al., 15 Jun 2026). The evaluation demonstrates infrastructure completeness—knowledge-to-skill transformation, lifecycle governance, and evolution—but does not yet systematically show downstream gains on large public agent benchmarks (Huang et al., 15 Jun 2026). Long-term stability is also open: continuous automated evolution raises issues of drift, compatibility, and regression (Huang et al., 15 Jun 2026).
A plausible implication is that SkillWiki marks a transition from isolated skill-learning methods toward a governed ecosystem in which skill assets are produced from heterogeneous evidence, linked to provenance, monitored in deployment, and revised under explicit lifecycle control. In that formulation, the distinctive contribution of SkillWiki is not a new skill representation alone, but the claim that knowledge, skills, and execution experience should co-evolve within a shared infrastructure (Huang et al., 15 Jun 2026).
