Awesome-Dynamic-Agent-Skills: Survey Evolving Agent Skills

Overview

Awesome-Dynamic-Agent-Skills is a research reading list accompanying a 2026 survey of skill libraries that change over time in LLM agent systems. It organizes work on how skills are represented, proposed, verified, retrieved, maintained, and governed, rather than treating agent skills as static instructions or tools.

The repository is most useful to researchers and practitioners who need a structured entry point into this literature or a shared vocabulary for comparing systems. Its project page and paper are linked from the repository README.

Key Features

• Groups surveyed papers by research area, including executable skill libraries, skill-aware reinforcement learning, infrastructure, benchmarks, and safety.
• Distinguishes six senses of “skill,” from executable code and natural-language procedures to parametric artifacts and boundary cases such as memories.
• Defines an eight-stage lifecycle spanning evidence collection, proposal, admission, storage, retrieval, maintenance, distillation, and governance.
• Provides a ten-operator vocabulary for describing library changes, including adding, refining, merging, pruning, composing, and reranking skills.
• Summarizes recurring empirical observations about verification, maintenance, retrieval scale, and skill abstraction.
• Links paper entries to external research sources and provides citation and contribution guidance.

Use Cases

• Researchers entering the dynamic-agent-skills field can use the categorized bibliography to find relevant papers and related surveys.
• Authors comparing skill-learning methods can map systems to a common lifecycle and operator vocabulary.
• Agent developers evaluating skill-library design can consult the reading list for work on verification, maintenance, retrieval, and governance.
• Security researchers can use the safety and audit section to locate studies of malicious skills, prompt injection, and supply-chain risks.

Getting Started

Browse the README for the taxonomy and paper list, or visit the project page. The repository is a reading resource and does not provide an installation command.

Alternatives

• agentskills: Covers the broader LLM skills ecosystem, including tools, frameworks, MCP, and computer-use agents, rather than focusing on dynamic skill libraries and their lifecycle.
• Awesome-Self-Evolving-Agents: Surveys agent self-evolution across models, environments, and applications; dynamic skill libraries are only one part of its broader scope.

Considerations

• This is a curated research resource, not an agent framework or a ready-to-run skill-management implementation.
• Paper summaries and reported findings are presented by the maintainers; consult the linked papers for methods, context, and evidence.
• The README describes a survey cutoff of May 31, 2026, so later research may not be covered.
• The repository reports one star and no forks in the supplied metadata, so community adoption appears limited.