🌟 Awesome Agent Skills 🌟

A curated paper list and resource hub for skill-centric LLM agent ecosystems.

Introduction · Taxonomy · Paper List · Benchmarks · Resources · Citation

Note

This repository accompanies our survey, A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications.

We collect papers, benchmarks, platforms, and ecosystem resources for understanding how reusable agent skills are represented, acquired, retrieved, selected, evolved, and governed.

If this repository is useful for your work, please consider starring it and citing the survey.

@article{zhou2026comprehensive,
  title   = {A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications},
  author  = {Zhou, Yingli and Wang, Shu and Su, Yaodong and Du, Wenchuan and Fang, Yixiang and Lin, Xuemin},
  journal = {arXiv preprint arXiv:2605.07358},
  year    = {2026}
}

🎯 Introduction

LLM-based agents are moving from passive response generation toward action-oriented task execution. They can call tools, retrieve memories, operate APIs, write code, and interact with external environments. However, tool access alone does not solve the harder procedural question: when should a capability be invoked, how should multiple steps be coordinated, how should failures be handled, and how should outputs be validated?

Our survey frames this bottleneck as the procedural gap. Agent skills address it by packaging reusable task-focused know-how into durable artifacts. Under this view, agents handle high-level intent interpretation, reasoning, and planning, while skills form the operational layer that makes execution reusable, inspectable, composable, and governable.

This repository tracks the emerging research landscape around agent skills, including:

Representative papers across the agent-skill lifecycle.
Benchmarks and evaluation protocols for skill-centric agents.
Platforms and repositories for discovering, sharing, and governing skills.
Application scenarios where reusable skills are becoming central to agent performance.

_{Growth of representative research on agent skills from April 2023 to April 2026.}

🧩 What Are Agent Skills?

In the survey, an agent skill is defined as a reusable procedural artifact with bounded scope. It externalizes task-focused know-how: not only what can be done, but also when to act, how to execute, what heuristics and failure modes matter, and how to judge completion.

Formally, a skill can be modeled as:

S = (M, R, C)

Component	Meaning	Examples
`M`	Main instruction document	A `SKILL.md`, SOP, checklist, or workflow prompt
`R`	Auxiliary resources	References, templates, helper scripts, notebooks, schemas
`C`	Applicability conditions	Trigger descriptions, metadata, dependencies, embeddings

Compared with raw tools, skills package the how-to layer around capability use. Compared with plain memory, skills are intended to be retrieved, executed, revised, and governed as reusable operational units.

🕰️ Historical Evolution of Skills

🗺️ Taxonomy

We organize the literature around the agent-skill lifecycle, following the survey's taxonomy.

Lifecycle Stage	Core Question	Representative Topics
Skill Representation	How is procedural know-how packaged?	Text-backed skills, code-backed skills, hybrid skills
Skill Acquisition	Where do skills come from?	Human-derived, experience-derived, task-derived, corpus-derived acquisition
Skill Retrieval & Selection	How does an agent choose the right skill at the right time?	Dense/sparse retrieval, generative retrieval, hierarchy/graph retrieval, context-aware routing, composition
Skill Evolution	How do skills improve safely over time?	Revision, validation, policy coupling, repository evolution, trust, rollback, deprecation

At a high level, skill-centric agent systems turn transient agent behavior into persistent capabilities:

experience / expertise / corpus / task
        ↓
skill acquisition
        ↓
skill representation
        ↓
retrieval + selection
        ↓
execution
        ↓
feedback, validation, evolution, governance

📄 Related Surveys

(OpenReview 2026) A Survey on Agent Skills: Externalized Procedural Knowledge in Language Models [Paper]
(Preprints 2026) A Survey of Agent Skills: Toward Procedural Infrastructure for LLM Agents [Paper]
(arXiv 2026) Agent Skill Evaluation and Evolution: Frameworks and Benchmarks [Paper]
(TechRxiv 2026) A Systematic Survey of Self-Evolving Agents: From Model-Centric to Environment-Driven Co-Evolution [Paper]
(arXiv 2026) Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering [Paper]
(arXiv 2026) SoK: Agentic Skills -- Beyond Tool Use in LLM Agents [Paper]

📑 Paper List

0. Foundations: Tools, Protocols, Retrieval, Memory

These works provide the infrastructure layer for agent skills: tool use, protocol-based capability access, retrieval, memory, and agentic execution loops.

(NeurIPS 2023) Toolformer [Paper]
(ICLR 2023) ReAct [Paper]
(arXiv 2023) HuggingGPT [Paper]
(arXiv 2023) ToolLLM [Paper]
(Anthropic 2024) Model Context Protocol (MCP) [Paper]
(OpenAI 2023) Function Calling [Paper]
(NeurIPS 2020) RAG [Paper]
(EMNLP 2020) DPR [Paper]
(arXiv 2025) In-depth Analysis of Graph-based RAG in a Unified Framework [Paper]
(arXiv 2025) ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation [Paper]
(arXiv 2025) Clue-RAG [Paper]
(arXiv 2025) BookRAG: A Hierarchical Structure-aware Index-based Approach for RAG on Complex Documents [Paper]
(arXiv 2025) EraRAG [Paper]
(arXiv 2023) MemGPT [Paper]
(arXiv 2023) Think-in-Memory [Paper]
(arXiv 2026) EverMemOS [Paper]
(arXiv 2026) HyperMem [Paper]
(arXiv 2026) MSA [Paper]

1. Skill Representation

Skill representation studies how reusable procedural artifacts are packaged for agents to load, inspect, execute, and maintain.

Text-Based Skills

(NeurIPS 2023) Reflexion [Paper]
(AAAI 2024) ExpeL [Paper]
(NeurIPS 2024) Buffer of Thoughts [Paper]
(arXiv 2026) Trace2Skill [Paper]
(arXiv 2026) Ctx2Skill [Paper] [Code]
(arXiv 2026) From Anatomy to Smells: An Empirical Study of SKILL.md in Agent Skills [Paper]

Code-Backed Skills

(NeurIPS 2023) Voyager [Paper]
(arXiv 2026) SkillCraft [Paper]
(ICLR 2026) PolySkill [Paper]
(arXiv 2025) Inducing Programmatic Skills for Agentic Tasks [Paper]
(arXiv 2026) Harnessing LLM Agents with Skill Programs [Paper]
(arXiv 2026) Skill-as-Pseudocode [Paper]

Hybrid Skills

(TPAMI 2025) JARVIS-1 [Paper]
(ICLR 2024) Synapse [Paper]
(arXiv 2025) SkillWeaver [Paper]
(arXiv 2026) AgentSkillOS [Paper]
(arXiv 2026) Agent Skills Should Go Beyond Text: The Case for Visual Skills [Paper]
(arXiv 2026) Skill-RM [Paper]
(arXiv 2026) AIP [Paper]

2. Skill Acquisition

Skill acquisition studies how new skills are constructed from human expertise, agent experience, task demands, or external corpora.

_{Overview of skill acquisition routes: human-derived, experience-derived, task-derived, and corpus-derived skills.}

Human-Derived

(arXiv 2026) SkillNet [Paper]
(arXiv 2026) AgentSkillOS [Paper]
(arXiv 2026) SoK: Agentic Skills [Paper]
(arXiv 2024) Agent Hospital [Paper]
(arXiv 2026) SciVisAgentSkills [Paper]

Experience-Derived

(NeurIPS 2023) Voyager [Paper]
(NeurIPS 2023) Reflexion [Paper]
(AAAI 2024) ExpeL [Paper]
(NeurIPS 2024) Buffer of Thoughts [Paper]
(arXiv 2026) Trace2Skill [Paper]
(ICML 2025) Agent Workflow Memory [Paper]
(arXiv 2025) AgentEvolver [Paper]
(arXiv 2025) G-memory [Paper]
(arXiv 2025) Nemori [Paper]
(arXiv 2026) PANDO [Paper]
(arXiv 2026) SkillGen: Verified Inference-Time Agent Skill Synthesis [Paper]

Task-Derived

(Findings EMNLP 2023) CREATOR [Paper]
(ICLR 2024) ToolMakers [Paper]
(arXiv 2024) Cradle [Paper]
(arXiv 2024) CodeAct [Paper]

Corpus-Derived

(arXiv 2023) AppAgent [Paper]
(arXiv 2024) AutoGuide [Paper]
(arXiv 2023) HuggingGPT [Paper]
(arXiv 2023) ToolLLM [Paper]
(arXiv 2024) DS-Agent [Paper]
(arXiv 2026) OpenSkill: Open-World Self-Evolution for LLM Agents [Paper]
(arXiv 2026) Ctx2Skill [Paper] [Code]

3. Skill Retrieval and Selection

Skill retrieval and selection ask how agents surface the right skill from a growing library, then decide whether to invoke, compose, or revise that skill under the current task state and budget.

_{Skill retrieval narrows the candidate space; skill selection decides what to execute, compose, or adapt.}

Retrieval

(ICLR 2025) ToolGen: Unified Tool Retrieval and Calling via Generation [Paper]
(arXiv 2023) ToolLLM [Paper]
(arXiv 2026) AgentSkillOS [Paper]
(arXiv 2026) SkillNet [Paper]
(arXiv 2026) GraphSkill [Paper]
(arXiv 2026) Skill Is Not Document [Paper]
(arXiv 2026) SkillDAG [Paper]

Selection and Routing

(arXiv 2024) AutoGuide [Paper]
(arXiv 2026) MemSkill [Paper]
(arXiv 2026) Memento-Skills [Paper]
(arXiv 2026) SkillRouter [Paper]
(arXiv 2026) GraSP [Paper]
(Findings ACL 2025) ToolExpNet [Paper]
(arXiv 2026) Maestro [Paper]
(arXiv 2026) Skill is Not One-Size-Fits-All [Paper]
(arXiv 2026) Skill or Skip? Learning Selective Skill Invocation in Agentic Tasks via Dual-Granularity Preference Learning [Paper]
(arXiv 2026) SkillsInjector [Paper]

4. Skill Evolution and Governance

Skill evolution studies how skills are revised, validated, optimized, synchronized, deprecated, and governed after deployment.

_{Skill evolution turns feedback, failures, validation, and governance signals into safer reusable capabilities.}

Skill Formation, Refinement, and RL Optimization

(arXiv 2024) TROVE [Paper]
(arXiv 2026) Memento-Skills [Paper]
(arXiv 2026) AutoSkill [Paper]
(arXiv 2026) EvoSkill [Paper]
(arXiv 2026) CoEvoSkills [Paper]
(arXiv 2026) Ctx2Skill [Paper] [Code]
(arXiv 2026) AutoRefine [Paper]
(arXiv 2026) SkillRL [Paper]
(arXiv 2026) Uni-Skill [Paper]
(arXiv 2026) ARISE [Paper]
(arXiv 2025) CASCADE [Paper]
(arXiv 2026) SkillOpt [Paper]
(arXiv 2026) MUSE-Autoskill [Paper]
(arXiv 2026) SkillRevise [Paper]
(arXiv 2026) SkillSmith [Paper]
(arXiv 2026) ReSkill [Paper]
(arXiv 2026) SkillComposer [Paper]
(arXiv 2026) SkillMaster: Toward Autonomous Skill Mastery in LLM Agents [Paper]
(arXiv 2026) SkillMAS: Skill Co-Evolution with LLM-based Multi-Agent System [Paper]
(arXiv 2026) COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows [Paper]
(arXiv 2026) MOCHA [Paper]
(arXiv 2026) SkillGrad [Paper]
(arXiv 2026) You Live More Than Once: Towards Hierarchical Skill Meta-Evolving [Paper]
(arXiv 2026) GRASP [Paper]
(arXiv 2026) FederatedSkill [Paper]
(arXiv 2026) Bayesian-Agent [Paper]
(arXiv 2026) SkillHone [Paper]
(arXiv 2026) SKILL0 [Paper]
(arXiv 2026) Skill1 [Paper]
(arXiv 2026) Skill0.5 [Paper]

Memory-Centric and Runtime Re-entry

(arXiv 2025) ReasoningBank [Paper]
(arXiv 2026) MemRL [Paper]
(arXiv 2026) MemSkill [Paper]
(arXiv 2025) MemEvolve [Paper]

Governance, Trust, and Ecosystem Risk

(arXiv 2026) SkillRouter [Paper]
(arXiv 2026) SkillNet [Paper]
(Preprints 2026) SkillOS [Paper]
(arXiv 2025) Audited Skill-Graph [Paper]
(Zenodo 2026) PoisonedSkills [Paper]
(arXiv 2026) Agent Skills: A Data-Driven Analysis of Claude Skills for Extending Large Language Model Functionality [Paper]
(arXiv 2026) Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis [Paper]
(arXiv 2026) SkillsVote [Paper]
(arXiv 2026) Library Drift [Paper]
(arXiv 2026) Harmless Yet Harmful [Paper]
(arXiv 2026) When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems [Paper]
(arXiv 2026) ClawHub Security Signals: When VirusTotal, Static Analysis, and SkillSpector Disagree [Paper]
(arXiv 2026) SkillHarm [Paper]
(arXiv 2026) SkillGuard [Paper]
(arXiv 2026) POISE [Paper]
(arXiv 2026) From Registry to Repository: How AI Agent Skills Are Written, Adapted, and Maintained [Paper]
(arXiv 2026) Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains [Paper]

🧪 Benchmarks and Evaluation

Skill-centric evaluation should measure more than final task success. Important dimensions include retrieval quality, selection utility, execution robustness, cost, recovery, transfer, and long-term library health.

AgentBench [Paper]
WebArena [Paper]
TaskBench [Paper]
STULIFE [Paper]
TRACE [Paper]
Evo-Memory [Paper]
SkillsBench [Paper]
SRA-Bench [Paper] [Code] [Dataset]
R3-Skill [Paper]
SkillHarm [Paper]
MalSkillBench [Paper]
SkillGenBench [Paper]
SkillEvolBench [Paper]
From Raw Experience to Skill Consumption [Paper]
When Skills Don't Help [Paper]
More Skills, Worse Agents? [Paper]
Benchmarking Security Risk Detection and Verification in Open Agentic Skill Ecosystems [Paper]
A Framework for Evaluating Agentic Skills at Scale [Paper]
SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use [Paper]

🛠️ Ecosystem Platforms and Resources

_{Growth of human-derived skills in emerging agent-skill platforms.}

Platform	Link	Focus
SkillNet	https://skillnet.openkg.cn/	Large-scale skill repository and organization
ClawHub	https://clawhub.ai/	Agent skill sharing and discovery
SkillHub	https://www.skillhub.club/	Community skill resources
SkillsMP	https://skillsmp.com/	Marketplace-style skill ecosystem
Skills.sh	https://skills.sh/	Agent skill publishing and reuse
BrowserAct Skills	https://github.com/browser-act/skills	Browser automation skills for AI agents with local Chrome reuse, isolated sessions, and human handoff
UIZZE	https://uizze.com	Free anti-UI-slop skill and 800,000+ real web and iOS product screens for Codex, Claude Code, Cursor, and other coding agents
Hermes Tweet	https://github.com/Xquik-dev/hermes-tweet	Hermes Agent X/Twitter plugin with bundled skill metadata and safe-default social workflow execution
Markstream Skills	https://github.com/Simon-He95/markstream-vue/tree/main/.agents/skills	Framework-specific skills for streaming Markdown across Vue, React, Svelte, Angular, Nuxt, Next.js, and Vue 2
ax	https://github.com/Necmttn/ax	Local telemetry graph for coding-agent sessions, skills, tools, and workflow recall
Tree Ring Memory	https://github.com/TerminallyLazy/Tree-Ring-Memory	Local-first memory framework for coding agents with lifecycle-aware recall, evidence-backed promotion, forgetting, and a portable Agent Skill package
UnifAPI Skills	https://github.com/unifapi-agent/skills	Public-data MCP and KOL pricing skill package for Codex and Claude-compatible agent workflows
RunAPI CLI Skill	https://github.com/runapi-ai/cli-skill	Agent workflow skill for running RunAPI image, video, music/audio, and LLM model jobs from Codex and Claude-compatible agents
Orkas VideoStudio	https://github.com/Orkas-AI/Orkas-VideoStudio	Source-installable video skills and MCP tools for Codex and Claude Code with editable timelines
Duvo Skills	https://github.com/duvoai/skills	Public agent skills for grocery and retail operations execution; SOP writer, run debugger, and CLI driver across existing systems
Before You Build Skill	https://github.com/bin1874/before-you-build-skill	Pre-build product and feature risk review skill for AI coding agents
skillZs	https://skillzs.dev/	Agent Skills discovery, guides, and security resources

🚀 Application Scenarios

_{Representative application scenarios where agent skills act as reusable operational units.}

💡 Research Opportunities

The survey highlights several open directions for future agent-skill research:

Unified skill schema: Common fields for scope, triggers, dependencies, versioning, resources, safety constraints, and provenance.
Resource-aware optimization: Jointly optimize retrieval, planning, execution, latency, tool cost, and risk.
Library evolution under non-stationarity: Handle API drift, changing task distributions, compatibility checks, rollback, and regression recovery.
Multimodal and domain-specific benchmarks: Evaluate skills in embodied, GUI, robotics, autonomous driving, UAV, healthcare, finance, and other constrained settings.
Causality-driven skill diagnosis: Attribute failures to retrieval mismatch, policy mis-selection, unsafe composition, stale dependencies, or tool malfunction.

🤝 Contributing

Contributions are welcome. If you want to add a paper, benchmark, project, or platform, please include:

Title
Venue and year
Link to paper, code, project, or website
Suggested category in this taxonomy

You can open an issue or submit a pull request directly.

Name		Name	Last commit message	Last commit date
assets		assets
README.md		README.md

Navigation Menu

Repository files navigation

🌟 Awesome Agent Skills 🌟

📚 Contents

🎯 Introduction

🧩 What Are Agent Skills?

🕰️ Historical Evolution of Skills

🗺️ Taxonomy

📄 Related Surveys

📑 Paper List

0. Foundations: Tools, Protocols, Retrieval, Memory

1. Skill Representation

Text-Based Skills

Code-Backed Skills

Hybrid Skills

2. Skill Acquisition

Human-Derived

Experience-Derived

Task-Derived

Corpus-Derived

3. Skill Retrieval and Selection

Retrieval

Selection and Routing

4. Skill Evolution and Governance

Skill Formation, Refinement, and RL Optimization

Memory-Centric and Runtime Re-entry

Governance, Trust, and Ecosystem Risk

🧪 Benchmarks and Evaluation

🛠️ Ecosystem Platforms and Resources

🚀 Application Scenarios

💡 Research Opportunities

🤝 Contributing

About

Releases

Packages

Contributors

Latest commit

History

Folders and files

Repository files navigation

🌟 Awesome Agent Skills 🌟

📚 Contents

🎯 Introduction

🧩 What Are Agent Skills?

🕰️ Historical Evolution of Skills

🗺️ Taxonomy

📄 Related Surveys

📑 Paper List

0. Foundations: Tools, Protocols, Retrieval, Memory

1. Skill Representation

Text-Based Skills

Code-Backed Skills

Hybrid Skills

2. Skill Acquisition

Human-Derived

Experience-Derived

Task-Derived

Corpus-Derived

3. Skill Retrieval and Selection

Retrieval

Selection and Routing

4. Skill Evolution and Governance

Skill Formation, Refinement, and RL Optimization

Memory-Centric and Runtime Re-entry

Governance, Trust, and Ecosystem Risk

🧪 Benchmarks and Evaluation

🛠️ Ecosystem Platforms and Resources

🚀 Application Scenarios

💡 Research Opportunities

🤝 Contributing

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors