This repository curates research papers on "agent skills" for LLMs, addressing the need for organized knowledge in this rapidly advancing field. It targets researchers and practitioners by providing a structured overview of techniques for skill learning, reasoning, orchestration, evaluation, and security, enabling rapid assessment of state-of-the-art methodologies.

How It Works

The project functions as a structured bibliography, categorizing seminal and recent papers on agent skills. It details each paper's title, authors, a concise summary of its core contribution, and a link. Organization spans key sub-domains: skill acquisition, reasoning, routing, benchmarking, and security, offering a thematic map of the research landscape.

Quick Start & Requirements

This is a curated list of research papers, not a runnable software project. No installation or execution commands or software prerequisites are provided.

Highlighted Details

• Skill Learning: Papers like SkillRL report state-of-the-art results on benchmarks (e.g., ALFWorld, WebShop), outperforming baselines by over 15.3%. Programmatic Skill Networks (PSN) show efficiency gains on Minecraft tasks.
• Orchestration & Ecosystems: SkillNet lists over 200,000 AI skills, improving agent rewards (+40%) and reducing steps (-30%). SkillOrchestra achieves superior routing performance while drastically cutting learning costs.
• Benchmarking & Evaluation: SkillsBench provides a framework where curated skills improve agent pass rates by an average of 16.2 pp. The Tool Decathlon benchmarks agents across 32 applications and 604 tools for long-horizon tasks.
• Security: Empirical studies reveal risks: 26.1% of over 42,000 audited skills had vulnerabilities. Agent skills are identified as a novel surface for prompt injection attacks.

Maintenance & Community

No information regarding maintenance, contributors, community channels, or roadmaps is provided.

Licensing & Compatibility

No specific software license is mentioned for the repository. The licensing of individual papers or associated codebases is not detailed.

Limitations & Caveats

As a curated list, this repository offers purely informational value as a research guide. Direct functionality or code is absent. Specific adoption blockers or technical requirements reside within the individual papers referenced, not this collection.