MMCPMarkets
WA

weishuyue/awesome-agentic-skills-research-papers

Developer tools
10 stars Quality 53 Trend 28NOASSERTION
Repository Install

A curated academic reading list on agentic skills.

Overview

A curated academic reading list on agentic skills. This repository is research-first. It is intentionally different from implementation catalogs such as awesome-agent-skills repositories that list ready-to-install SKILL.md packages. - Papers where agent skills, skill libraries, skill routing, skill distillation, skill evolution, or skill evaluation are a primary object of study. - Closely related precursor papers on tool learning, experiential learning, and executable skill libraries. - Surveys, benchmarks, datasets, and security/governance papers that help define the field. - General LLM agent papers unless they introduce or evaluate reusable skills. - Marketplaces or skill-package collections unless they are used as research artifacts. - Prompt libraries with no evaluation or research contribution. No standalone security-only paper is listed yet.

README

🌟 Awesome Agentic Skills Research Papers

A curated academic reading list on agentic skills.

Last checked: 2026-11-02.

Latest literature review: July–November 2, 2026. Official published versions are preferred where verified; Findings, Industry Track, pre-proceedings, and arXiv preprints are labeled separately.

🗺️ Scope

This repository is research-first. It is intentionally different from implementation catalogs such as awesome-agent-skills repositories that list ready-to-install SKILL.md packages.

Included:

  • Papers where agent skills, skill libraries, skill routing, skill distillation, skill evolution, or skill evaluation are a primary object of study.
  • Closely related precursor papers on tool learning, experiential learning, and executable skill libraries.
  • Surveys, benchmarks, datasets, and security/governance papers that help define the field.

Not included by default:

  • General LLM agent papers unless they introduce or evaluate reusable skills.
  • Marketplaces or skill-package collections unless they are used as research artifacts.
  • Prompt libraries with no evaluation or research contribution.

📚 Contents

📄 Paper list

Surveys and Position Papers

Year Paper Links Why it matters
2026 Harnessing Agent Skills: Architectural Patterns and a Reference Architecture for Skill-Mediated LLM Agents arXiv Catalogue of 10 architectural patterns and a 4-layer reference architecture (Supply Chain, Mediation, Execution Control, Evidence & Feedback) for skill harnessing.
2026 SoK: Agentic Skills – Beyond Tool Use in LLM Agents arXiv Frames skills as reusable procedural capabilities beyond atomic tool calls, with lifecycle, representation, scope, and security taxonomies.
2026 Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward arXiv Focuses on the SKILL.md/progressive-disclosure abstraction, MCP relationship, acquisition, deployment, and trust tiers.
2026 A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications arXiv Organizes the literature around representation, acquisition, retrieval, and evolution.
2025 LLM-Based Agents for Tool Learning: A Survey Springer Useful bridge from tool learning to skill-equipped agents.
2024 A Review of Prominent Paradigms for LLM-Based Agents: Tool Use, Planning, and Feedback Learning arXiv Provides the broader agent paradigm background for skill-based systems.
2023 A Survey on Large Language Model based Autonomous Agents arXiv General LLM-agent survey; useful for positioning skills against memory, planning, and action modules.
2023 The Rise and Potential of Large Language Model Based Agents: A Survey arXiv Broad overview of single-agent, multi-agent, and human-agent scenarios.
2026 Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries TMLR 2026 Organizes evolving skill libraries through a skill taxonomy, an eight-stage lifecycle, and verification, maintenance, and governance requirements.

Foundations and Precursors

Year Paper Links Why it matters
2023 Toolformer: Language Models Can Teach Themselves to Use Tools arXiv Early self-supervised tool-use training; a precursor to learning callable procedural capability.
2023 API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs arXiv Benchmark and dataset for tool-augmented LLM behavior.
2023 On the Tool Manipulation Capability of Open-source Large Language Models arXiv Studies practical tool manipulation and creates a ToolBench benchmark.
2023 ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs arXiv Large-scale API/tool learning framework and dataset.
2023 Voyager: An Open-Ended Embodied Agent with Large Language Models arXiv Introduces an ever-growing executable code skill library for Minecraft agents.
2023 ExpeL: LLM Agents Are Experiential Learners arXiv Learns from task experience without model-weight updates; important precursor to skill distillation.
2024 SkillAct: Using Skill Abstractions Improves LLM Agents OpenReview Shows that prompting with reusable skill abstractions improves interactive task performance.
2024 LLMs in the Imaginarium: Tool Learning through Simulated Trial and Error arXiv Uses simulated trial-and-error and memory to learn tool use behavior.

Skill Acquisition and Self-Improvement

Year Paper Links Why it matters
2024 Agentic Skill Discovery arXiv LLM-driven discovery of robotic skills from zero initial skill library.
2026 Reinforcement Learning for Self-Improving Agent with Skill Library ACL 2026 Introduces SAGE, using Skill Augmented GRPO and sequential rollouts to accumulate reusable skills.
2025 CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution arXiv Studies cumulative skill creation and evolution for AI-assisted scientific workflows.
2026 SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning arXiv Builds a hierarchical SkillBank from experience and co-evolves skill library and policy.
2026 Agentic Proposing: Enhancing Large Language Model Reasoning via Compositional Skill Synthesis ICML 2026 Uses modular reasoning skills to synthesize verifiable training problems.
2026 AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution arXiv Derives, maintains, and reuses skills from dialogue and interaction traces.
2026 Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills EMNLP 2026 Distills trajectory-local lessons into transferable agent skills.
2026 Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents arXiv Converts agent trajectories into dynamic training-only skill supervision.
2026 SkillX: Automatically Constructing Skill Knowledge Bases for Agents arXiv Builds plug-and-play hierarchical skill knowledge bases from trajectories.
2026 MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction arXiv Automatically induces generalizable skills from successful trajectories with quality guarantees.
2026 Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning arXiv Jointly optimizes skill selection, utilization, and distillation from a shared task-outcome signal.
2026 SkillGraph: Skill-Augmented Reinforcement Learning for Agents via Evolving Skill Graphs arXiv Represents skills as graph nodes with prerequisite, enhancement, and co-occurrence edges.
2026 SkillMaster: Toward Autonomous Skill Mastery in LLM Agents arXiv Trains agents to create, refine, and select their own accumulated skills.
2026 SkillEvolver: Skill Learning as a Meta-Skill arXiv Treats skill learning itself as a reusable meta-skill that authors, deploys, and refines skills.
2026 SkillGen: Verified Inference-Time Agent Skill Synthesis arXiv Synthesizes auditable skills from success and failure trajectories, then verifies net intervention effect.
2026 SkillGrad: Optimizing Agent Skills Like Gradient Descent arXiv Optimizes skill packages with trajectory-level loss evidence and text-based gradients.
2026 COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation arXiv Distills person- or role-grounded traces into inspectable, correctable skill packages.
2026 OpenSkill: Open-World Self-Evolution for LLM Agents arXiv Builds skills and verification signals from scratch in open worlds with no target-task supervision; demonstrates strong transferability across models.
2026 Ctx2Skill: From Context to Skills: Can Language Models Learn from Context Skillfully? arXiv Self-evolving multi-agent self-play framework that discovers, refines, and selects context-specific skills without human supervision.
2026 SoftSkill: Behavioral Compression for Contextual Adaptation arXiv Compresses long textual skills into a compact continuous prefix, enabling efficient behavioral adaptation and significant context reduction while keeping the base model frozen.
2026 SkillOpt: Executive Strategy for Self-Evolving Agent Skills arXiv Train agent skills like you train neural networks — with epochs, (mini-)batchsize, learning rates, and validation gates — but without touching model weights.
2026 CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification arXiv Co-evolves a Skill Generator and a Surrogate Verifier to autonomously construct and iteratively refine complex multi-file skill packages without ground-truth test supervision.
2026 Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents ICML 2026 Learns reusable procedural skills without weight updates, with explicit activation and termination conditions, verification gates, and library maintenance.
2026 Meta Context Engineering via Agentic Skill Evolution ICML 2026 Co-evolves context-engineering skills and context artifacts through a meta-level agent and execution feedback.
2026 Harnessing LLM Agents with Skill Programs EMNLP 2026 Turns advisory skills into executable Program Functions that intervene in the agent loop and support validated evolution.
2026 SkillPyramid: A Hierarchical Skill Consolidation Framework for Self-Evolving Agents EMNLP 2026 Consolidates skills in a hierarchy and composes, validates, and incorporates new skills during execution.
2026 Ability-Aligned Skill Evolution for LLM Agents via Reinforcement Learning Findings of EMNLP 2026 Aligns skill-bank evolution with agent competence using failure attribution, verified updates, and per-task-type rollback.
2026 SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization Findings of EMNLP 2026 Uses a curriculum that progressively withdraws skill context during RL to internalize skills without runtime retrieval.
2026 SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents Findings of EMNLP 2026 Jointly trains action and skill-invocation decisions while verifying and refining the skill bank through environment interaction.
2026 MMG2Skill: Can Agents Distill In-the-Wild Guides into Self-Evolving Skills? Findings of EMNLP 2026 Studies guide-to-skill learning with a benchmark and editable skills revised from trajectory-level failure analysis.
2026 RepairGate: RL-Trained Skill Repair and Abstention for Lifelong Agent Libraries Findings of EMNLP 2026 Trains a post-trial gate to accept, block, or override skill patches while controlling library growth.
2026 EVISKILL: Grounding Skill Evolution in Replayable Evidence arXiv preprint Links skill edits to replayable evidence and uses targeted re-execution and global validation before admission.
2026 Evolving to Match Capabilities: Improving Harness Self-Evolution via Skill Adaptation arXiv preprint Studies why skills transfer unevenly across model capabilities and uses a meta-skill to adapt guidance to the executor.
2026 SkillAlchemy: Open-World Agent Skill Creation arXiv preprint / Code Creates source-grounded skills from underspecified briefs by discovering implicit requirements, admitting procedures within evidence-supported scope, and compiling structured skill packages.
2026 MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation arXiv Frames skills as long-lived assets in a creation, memory, management, evaluation, and refinement lifecycle.
2026 SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision arXiv Refines imperfect LLM-authored skills using execution traces, repair principles, re-execution, and empirical utility selection.

Skill Retrieval, Routing, and Context Construction

Year Paper Links Why it matters
2026 SkillRouter: Skill Routing for LLM Agents at Scale COLM 2026 pre-proceedings Studies large-scale skill routing and shows full skill bodies are crucial routing signal.
2026 Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills EMNLP 2026 Retrieves dependency-aware skill bundles instead of isolated semantically similar skills.
2026 SkillsInjector: Dynamic Skill Context Construction for LLM Agents arXiv Learns adaptive skill budgets and set-aware rendering for injected context.
2026 The Scaling Laws of Skills in LLM Agent Systems arXiv Empirically studies how routing and execution degrade or recover as skill libraries scale.
2026 Skill Retrieval Augmentation for Agentic AI Findings of EMNLP 2026 Defines retrieval-based skill augmentation and SRA-Bench to evaluate retrieval, incorporation, and application separately.
2026 SkillAlign: Aligning Skill Interfaces for LLM-based Agents EMNLP 2026 Evaluates how alternative skill exposure interfaces affect utility and context cost with candidate skills held fixed.
2026 SKILL.state: Scalable Long-Horizon Agent Skills EMNLP 2026 Industry Track Executes procedural skills with explicit mutable state instead of an ever-growing conversation history.

Benchmarks and Evaluation

Year Paper Links Why it matters
2026 How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings arXiv Benchmarks skill utility under realistic retrieval and refinement conditions.
2026 SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents arXiv Provides a retrieval benchmark with public agent skills, taxonomy, training samples, and evaluation queries.
2026 SkillMOO: Multi-Objective Optimization of Agent Skills for Software Engineering arXiv Optimizes SE skill bundles for both pass rate and inference cost.
2026 ClawTrace: Cost-Aware Tracing for LLM Agent Skill Distillation arXiv Adds cost-aware trace cards and skill patches for distillation pipelines.
2026 OpenSkillEval: Automatically Auditing the Open Skill Ecosystem for LLM Agents Findings of EMNLP 2026 Constructs artifact-grounded tasks to compare open skills and skill-augmented agents across five applications.
2026 Skill Following: Evaluating Actual Skill Use in Retrieval-Enabled LLM Agents Findings of EMNLP 2026 Uses matched skill-enabled and skill-disabled runs on retrieved tasks to distinguish aggregate retrieval lift from actual skill utility.
2026 GitSkills: A Dataset of Agent Skills on GitHub arXiv preprint Provides a large-scale corpus of SKILL.md occurrences with content hashes and metadata for reuse, maintenance, and security research.
2026 Demystifying Agent Skills: Why They Work-Until They Don’t arXiv preprint Uses controlled comparisons and paired trajectory analysis to study skill representations, procedural anchoring, and retrieval bottlenecks.
2026 Agent Skill Evolution: How Revisions Affect Coding Agents arXiv preprint Studies real skill revisions and measures changes in rule compliance, agent behavior, correctness, and context cost.
2026 SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents arXiv Treats skill generation itself as the benchmark target, covering task-conditioned and task-agnostic generation from repository- and document-grounded sources.
2026 SkillLearnBench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real-World Tasks arXiv Evaluates continual skill generation methods across skill quality, execution trajectory, and task outcome, including skill-creator-style baselines.

Application-Specific Skill Systems

Year Paper Links Why it matters
2026 CUA-Skill: Develop Skills for Computer Using Agent arXiv Builds a structured skill base for computer-using agents on Windows applications.
2026 SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support arXiv Creates and refines domain-specific enterprise support skills from knowledge bases and operational failures.
2026 PANDO: Efficient Multimodal AI Agents via Online Skill Distillation arXiv Online skill distillation for multimodal web agents with efficiency metrics.
2026 Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report arXiv Uses harness-state augmentation to train a live-commerce agent to adapt to revised reusable skills within a changing runtime.
2026 Efficient Skill Grounding via Code Refactoring with Small Language Models ICML 2026 Grounds reusable executable skills across embodiments by refactoring execution bindings while preserving skill control structure.
2026 SkillMe: Evolving Agent Skills for Recurring Personal Workflows EMNLP 2026 Adapts reusable skills through repeated execution in recurring personal workflows with local tools and formats.
2026 SCAFFOLD: Self-Improving Web Agents via Recursive Parametric Skill Abstraction Findings of EMNLP 2026 Builds recursive executable skill hierarchies for web agents, compacts redundant skills, and distills skill-guided trajectories.
2026 SkillForge: Compositional Skill Synthesis with Verification-in-the-Loop for Generating Formally Verified Dafny Programs Findings of EMNLP 2026 Composes atomic skills for specification, synthesis, invariant generation, and repair with decidable verification.

Security and Governance

Year Paper Links Why it matters
2026 SkillTrojan: Backdoor Attacks on Skill-Based Agent Systems ICML 2026 Studies backdoors in skill implementations and composition, with a dataset for evaluating skill-level attacks.

For additional security and governance angles in the current list, see:

🔗 Related Awesome Lists

These are useful, but have a different center of gravity from this repository.

🤝 Contributing

If you want to add a paper, benchmark, project, or resource, please include:

  1. Title
  2. Venue and year
  3. Link to paper, code, or website
  4. Suggested category in this taxonomy

Suggested entry format:

| 2026 | Paper Title | [arXiv](https://arxiv.org/abs/xxxx.xxxxx) / [Code](https://github.com/...) | One-sentence reason this belongs in agentic skill research. |

Inclusion checklist:

  • The paper treats skills, reusable procedures, tool-using routines, executable behaviors, skill libraries, or skill routing as a main contribution.
  • The link points to a primary source when possible: arXiv, ACL Anthology, OpenReview, publisher page, project page, or official repository.
  • Prefer the official published version and its title/year when available; label Findings, Industry Track, workshops, pre-proceedings, and preprints accurately. Keep one entry per work rather than listing each version separately.
  • The note explains the agentic-skill relevance, not just the general LLM-agent relevance.
View the full README on GitHub

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add weishuyue/awesome-agentic-skills-research-papers
New to MCP? Start with the mcpskill.org guide →