SkillFab: An Agent-Native Skill Production Platform

Jul 16, 2026

Upload by Ivy

SkillFab is an agent-native platform for transforming missing capabilities into reviewed, reusable Agent Skills through a demand-first collaboration workflow. At runtime, agents search for existing skills; if no adequate skill exists, the unmet capability becomes a first-class issue that triggers repository-backed submission. Development uses Git-style commit evidence, maintainer review, and registry publication, while exposing the same lifecycle via web, REST, and MCP for shared state. The system supports long-running reviewability and recovery, and demonstrates OS-detect, Docker-to-skill, and external optimization cases.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages12

Reading Time30 min

Introduction

Platform Model

System Architecture

Implementation Surfaces (Web/REST/MCP)

Workflow State and Event Histories

Case Studies

OS-Detect Skill Run

Docker Research Package to Skill Knowledge

SkillOpt External Optimization Lifecycle

arXiv :2607 .03780v 1 [ cs . SE] 4 Jul 2026 SkillFab: An Agent-Native Skill Production Platform System Design Technical Report Anjie Xu 1,4 , Yifeng Cai 1 , Yi Li2 , Zixing Wang3 , Zhiyu Zhang 1 , Jingfan Chen 1 , Ruohan Xu 1 , Leye Wang 1 ∗ 1 Peking University 2 Huawei 3 Northwestern Polytechnical University 4 Zhongguancun Academy Abstract SkillFab is an agent-native platform for turning missing capabilities into reviewed, reusable Agent Skills. At runtime, agents first search for reusable skills; when no adequate skill exists, the unmet capability becomes a demand-first issue before any repository or implementation branch needs to exist. Development then proceeds through a SkillFab-managed repository, Git-ingested commit evidence, maintainer review, and registry publication. The same lifecycle is exposed through web, REST, and MCP surfaces, so humans, scripts, and external agents operate on shared state rather than separate task logs. The current system uses scoped Git push URLs, native range commit ingestion, workflow-state reads, and workflow-event histories to make long-running agent work reviewable and recoverable. We document the platform model, architecture, implemented capabilities, and three case studies: an end-to-end OS-detect skill run, a Docker research package that converts operational practice into reusable skill knowledge, and an external optimization case showing how improved skill artifacts can enter SkillFab as reviewable, versioned submissions. Deployment: [skillfab.ai](skillfab.ai) 1 Introduction Modern agents can solve individual tasks with tool use, reflection, and environment interaction, but successful procedures rarely become maintained artifacts. A later agent may need the same capability and still receive only a prompt fragment, a transcript, or no reusable object. Capability growth then depends on rediscovery and informal copying rather than ownership, provenance, and revision. This gap appears across work on reasoning-and-acting agents, self-supervised tool use, verbal self-improvement, explicit skill accumulation, and agent benchmarks.[1, 2 , 3 , 4 , 5 , 6 , 7] SkillFab makes skill production a collaboration workflow. A missing or inadequate skill becomes an issue even when no repository, package, or prior skill exists. Implementation happens in a repository-backed submission, where [SKILL.md](SKILL.md) and supporting files are tied to ingested commits. Maintainers review the package snapshot, publish accepted versions to the registry, and let later reuse open new issues when the skill fails or becomes incomplete. This connects agent skill accumulation to established software-engineering concerns around reuse, collaborative production, review, and maintenance.[8, 9 , 10] SkillFab resembles GitHub in vocabulary but not in starting point. GitHub issues are usually scoped by an existing repository or project;[11, 12 , 13] SkillFab issues can name an unmet agent ∗ Corresponding author. capability before code exists. The design contribution is the coupling of Git-style software evidence[14, 15] with agent-native execution surfaces, skill-specific review and certification, registry retrieval, and workflow-state recovery. The implementation described here is running code: a Hono application exposes web, REST, and MCP surfaces;[16] SQLite stores workflow state;[17] a separate git-server accepts scoped pushes; and a native Rust addon ingests commit ranges and owns many transactional paths. Broader evaluation, trust policy design, continued operational hardening, and larger deployments remain ongoing work. The public SkillFab service is available at [https://skillfab.ai](https://skillfab.ai), which serves as the project website and live system entry point.[18] SkillTester is the twin evaluation project: its public service is available at [https://skilltester.ai](https://skilltester.ai), and its technical report focuses on benchmarking agent skills for utility and securi

FAQs

What problem does SkillFab address for agent systems?

SkillFab addresses the gap where newly discovered agent procedures are not turned into maintained, reusable artifacts, leaving later agents with only prompt fragments or transcripts instead of owned, reviewable skills.

How does SkillFab handle a missing capability before any code or repository exists?

An unmet capability becomes a first-class issue in a demand-first model, which can exist without a repository or prior published skill, and only then triggers implementation via a repository-backed submission workflow.

How does SkillFab enable reviewability and reuse of long-running agent work?

SkillFab ties submissions to ingested Git commit evidence and review packets, publishes accepted versions to a registry, and uses workflow-state storage plus workflow-event histories to make execution recoverable.

Jul 16, 2026

Upload by Ivy