How SkillNet Turns AI Agent Experience into Reusable Skills

Background and Motivation

Current AI agents can call tools and perform complex tasks, but they lack a systematic way to accumulate and transfer skills, leading to repeated "reinventing the wheel" across different scenarios. Researchers from Zhejiang University, Alibaba, Ant Group, and Tencent have organized more than 200k skills into a structured graph to address this gap.

SkillNet Overview

SkillNet introduces a complete skill‑lifecycle framework consisting of three core capabilities: creation , evaluation , and connection . The system aims to convert fragmented agent experiences into standardized, reusable skill packages.

1. Skill Creation – Extracting Abilities from Heterogeneous Sources

SkillNet automatically generates skills from various inputs:

Execution traces : logs of agent actions.

GitHub projects : open‑source code repositories.

Office documents : PDFs, PPTs, Word files.

Natural‑language prompts : user‑provided descriptions.

An LLM‑driven pipeline structures these raw materials into standardized Skill Packages that contain metadata, execution commands, and optional code resources.

https://export.arxiv.org/pdf/2603.04448
SkillNet: Create, Evaluate, and Connect AI Skills
http://skillnet.openkg.cn
https://github.com/zjunlp/SkillNet

2. Skill Evaluation – Five‑Dimensional Quality Assurance

Unlike platforms that rely solely on community likes or download counts, SkillNet implements an automated, multi‑dimensional evaluation framework covering:

Safety : guards against dangerous operations and prompt injection attacks.

Integrity : checks whether steps are complete and dependencies are explicit.

Executability : verifies that the skill runs successfully in a sandbox.

Maintainability : assesses modularity and backward compatibility.

Cost Awareness : measures time, compute, and API overhead.

3. Skill Connection – Building an Inferable Skill Graph

SkillNet organizes skills using a three‑layer ontology:

Classification layer : ten high‑level domains such as Development, AIGC, Science.

Relation layer : defines relationships like similar_to, compose_with, depend_on, belong_to.

Package layer : groups related skills into task‑oriented collections.

Experimental Results

SkillNet was evaluated on three benchmarks—ALFWorld (home environment), WebShop (online shopping), and ScienceWorld (scientific experiments). The results show:

Average reward increase of roughly 40%.

Number of execution steps reduced by about 30%.

Consistent gains across models ranging from lightweight o4 Mini to powerful Gemini 2.5 Pro.

Strong generalization to unseen tasks.

Conclusion

SkillNet demonstrates that turning agent experience into a structured, evaluable, and connectable asset dramatically improves efficiency and robustness. By standardizing creation, providing multi‑dimensional evaluation, and linking skills through a graph, it enables a future where a single expert can orchestrate a society of agents, and skills become the core medium for knowledge exchange and capability accumulation.