Current AI agents can flexibly invoke tools and execute complex tasks, yet their long-term advancement is hindered by the lack of systematic accumulation and transfer of skills. Without a unified mechanism for skill consolidation, agents frequently ``reinvent the wheel'', rediscovering solutions in isolated contexts without leveraging prior strategies. To address this challenge, we introduce SkillNet, an open infrastructure for creating, evaluating, and organizing AI skills at scale. SkillNet structures skills within a unified ontology that supports creating skills from heterogeneous sources, establishing rich relational connections, and performing multi-dimensional evaluation across Safety, Completeness, Executability, Maintainability, and Cost-awareness. Our infrastructure integrates a repository of over 600,000 skills, an interactive platform, and a versatile Python toolkit. Experiments on ALFWorld, WebShop, and ScienceWorld show 40% higher average rewards and 30% fewer execution steps across multiple backbone models. Furthermore, SkillNet-Gym benchmarks skill retrieval, utilization, and composition, while SkillNet-Fabric enables task-specific skill routing through lightweight Wikis. By formalizing skills as evolving, composable assets, SkillNet provides a robust foundation for agents to move from transient experience to durable mastery.
Figures & tables
Figure 1: Overview of SkillNet . SkillNet organizes large-scale agent skills into a structured skill network, modeling rich relations (e.g., similarity, composition, and dependency), supporting multi-dimensional evaluation, and providing unified interfaces for skill discovery, creation, and analysis.
Figure 2: End-to-end Pipeline of SkillNet. SkillNet transforms heterogeneous user inputs and open internet resources into executable skills through automated skill creation and multi-dimensional evaluation, and organizes high-quality skills into a structured network to support search, download, analysis, and contribution.
Figure 3: The Skill Ontology for SkillNet. It consists of three levels: the Skill Taxonomy (top) defines functional categories; the Skill Relation Graph (middle) models inter-skill dependencies and semantic associations; and the Skill Package Library (bottom) organizes skills into modular, task-oriented bundles.
Figure 4: Multi-dimensional Skill Evaluation and Reliability Analysis. (a) Quality distribution of curated skills in SkillNet, assessed using a three-level grading scheme, where most skills are rated Good or Average . (b) Reliability validation of the automated evaluator on 200 randomly sampled skills. The heatmap shows mean absolute error (MAE) and quadratic weighted kappa (QWK) between human annotators and the model evaluator, with near-perfect QWK and low MAE across all dimensions, demonstrating the robustness and scalability of the evaluation framework.
Figure 5: Practical use examples of skillnet-ai. The package offers a unified functional experience through both a command-line interface (left) for interactive usage and a Python library (right) for seamless integration into AI development pipelines.
Model
Method
ALFWorld
WebShop
ScienceWorld
Seen
Unseen
R ↑
S ↓
Seen
Unseen
R ↑
S ↓
R ↑
S ↓
R ↑
S ↓
R ↑
S ↓
DeepSeek V3.2
React
66.43
19.51
69.40
19.27
31.55
24.06
69.86
17.59
64.67
19.26
Expel
67.86
18.86
76.12
17.41
29.23
24.00
74.91
15.98
74.09
17.53
+ SkillNet
80.60
14.54
83.57
14.81
46.18
17.84
84.87
11.89
81.31
12.48
Gemini 2.5 Pro
React
60.00
18.72
61.94
19.18
31.66
22.12
58.24
18.42
56.13
19.07
Table 1: Experimental results on ALFWorld, WebShop, and ScienceWorld. R denotes average reward, S denotes average steps (↑ indicates the larger values are better, and ↓ denotes the smaller values are better). The top-performing results are highlighted in bold , and the second-best are underlined .
Figure 6: Performance comparison across diverse methods and models. The results illustrate that SkillNet consistently outperforms React and Few-shot baselines, achieving significantly higher average rewards (top) and reduced average steps (bottom) across ALFWorld, WebShop, and ScienceWorld.
Figure 7: The overview of SkillNet-Gym construction pipeline.
Benchmark
Benchmark Construction
Benchmark Scope
Evaluation Focus
Real-World Source
Automatic Construction
Dynamic Extensibility
Domain Coverage
Skill Pool
Task Numbers
Skill Construction
Skill Composition
SkillsBench [ 38 ]
✓ –
✗
✗
✓
✗
87
✗
✗
SWE-Skills-Bench [ 39 ]
✓ –
✗
✗
✗
✗
565
✗
✗
SkillLearnBench [ 40 ]
✓ –
✗
✗
✓ –
✗
20
✓
✗
WildSkillBench [ 41 ]
✓ –
✗
✗
✓
✓
87
✓
✓
SkillRouter [ 42 ]
✓
✓
✓
✓
✓
75
✗
✓
Table 2: Comparison of existing agentic skills benchmarks across three dimensions: Benchmark Construction , Benchmark Scope and Evaluation Focus . “ ✓ ” indicates fully coverage, “ ✓ – ” indicates partially coverage, and “ ✗ ” indicates not coverage. Detailed explanations for each trait are provided in Appendix 12 .
Table 3: Comparison across different harnesses and models under different skill usage settings. No Skill uses no external skills; Official Skill directly provides the task-associated official skills; Skill-Creator evaluates execution with skills constructed from upstream resources; Skills Retrieve requires agents to retrieve relevant skills from a large wild skill pool before execution; Skills Orchestration additionally requires dependency-aware organization of the selected skills into an execution workflow. Best and second-best results within each column are shown in bold and underlined, respectively.
Recent AI agents can flexibly invoke skills to solve complex tasks, but their long-term improvement is fundamentally constrained by a lack of systematic skill construction, accumulation, and transfer. In particular, without a unified framework for skill consolidation, agents tend to redundantly construct similar capabilities across different tasks, are unable to effectively transform experience into reusable assets, and struggle to generalize task-specific skills to novel scenarios. To address this limitation, we propose SkillPyramid, a skill consolidation framework that reuses existing skill experience for broader task generalization. Operating on a hierarchical skill topology, SkillPyramid further introduces a self-evolution mechanism that enables agents to compose, validate, and incorporate new skills during task execution. Experiments on ALFWorld, WebShop, and ScienceWorld across four backbone models show that SkillPyramid substantially increases the average reward by 38.0% and reduces execution steps by 27.7%. Overall, our method transforms a skill collection from a static resource pool into a dynamic evolution system.
Yuan Xiong, Ziqi Miao, Qian Chen +5
1The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · 3Shanghai Artificial Intelligence Laboratory, Shanghai, China · 4Beijing Academy of Artificial Intelligence, Beijing, China +1
Skills are a promising way to improve LLM agent capabilities without retraining, while keeping the added procedure reusable and controllable. However, high-quality skills are still largely written by hand. We introduce SkillGen, a multi-agent framework that synthesizes a single auditable skill from trajectories generated by a base agent. The output is a human-readable artifact that can be inspected before use. Rather than merely summarizing trajectories, SkillGen leverages contrastive induction over both successful and failed trajectories to identify reusable success patterns, recurring failure modes, and behaviors that appear in nearby successes but are missing from failures. SkillGen then generates candidate skills and iteratively refines the skill. A key novelty in SkillGen is that we model agent skills as interventions to empirically verify the net effect of skills on the overall performance. Specifically, we compare outcomes on the same instances with and without the skill, so that we account for both repairs (cases where the skill fixes a baseline failure) and regressions (cases where the skill breaks a baseline success). Across a broad range of agents and datasets, SkillGen consistently improves held-out performance, outperforms existing skill-generation baselines, and produces skills that transfer across models.
Yuchen Ma, Yue Huang, Han Bao +5
Munich Center for Machine Learning, LMU Munich · University of Notre Dame · Microsoft Research
Learning from experience is critical for building capable large language model (LLM) agents, yet prevailing self-evolving paradigms remain inefficient: agents learn in isolation, repeatedly rediscover similar behaviors from limited experience, resulting in redundant exploration and poor generalization. To address this problem, we propose SkillX, a fully automated framework for constructing a plug-and-play skill knowledge base that can be reused across agents and environments. SkillX operates through a fully automated pipeline built on three synergistic innovations: (i) Multi-Level Skills Design, which distills raw trajectories into three-tiered hierarchy of strategic plans, functional skills, and atomic skills; (ii) Iterative Skills Refinement, which automatically revises skills based on execution feedback to continuously improve library quality; and (iii) Exploratory Skills Expansion, which proactively generates and validates novel skills to expand coverage beyond seed training data. Using a strong backbone agent (GLM-4.6), we automatically build a reusable skill library and evaluate its transferability on challenging long-horizon, user-interactive benchmarks, including AppWorld, BFCL-v3, and τ2-Bench. Experiments show that SkillKB consistently improves task success and execution efficiency when plugged into weaker base agents, highlighting the importance of structured, hierarchical experience representations for generalizable agent learning. Our code will be publicly available soon at https://github.com/zjunlp/SkillX.
Chenxi Wang, Zhuoyun Yu, Xin Xie +8
1Zhejiang University · 2Ant Digital Technologies, Ant Group.