Agent Skills Survey

Abstract

As large language model (LLM)-powered autonomous agents increasingly demonstrate strong capabilities in complex real-world tasks, recent research has focused on leveraging skills to endow these agents with domain-specific expertise and adaptive competence. In this survey, we present the first systematic overview of LLM-powered Agent Skills from the perspective of procedural memory, conceptualizing skills as a core form of procedural knowledge that enables agents to internalize, retain, and reuse task-solving procedures through interaction and experience. We comprehensively examine what Agent Skills are, why they are essential for agentic intelligence, and how they can be effectively acquired, represented, invoked, and refined within LLM-based systems. Furthermore, we review representative applications across diverse domains and discuss key challenges and future opportunities in developing and generalizing Agent Skills.

Key Words: Agent Skills, Agentic AI, LLM Agents, Procedural Memory

Agent Skills for Problem Solving

Figure: Agent Skills for Problem Solving.

Skill Acquisition: Through repeated task interactions, agents convert recurring action patterns into reusable skills that encode experience-driven procedural knowledge.
Skill Representation: Skills are stored as modular, discoverable units of procedural memory with lightweight metadata, exposing details only when needed.
Skill Invocation: When a task fits a skill's scope, the agent retrieves and executes that skill (or related tools) to run complex workflows reliably.
Skill Refinement: Skills are continuously updated from outcomes and feedback to fix errors, improve efficiency, and adapt to new contexts.

Anthropic Agent Skills

Agent Skills Implementation

Method	Acquisition	Representation	Invocation	Refinement	Resource Link
Agent Skills	Manual	File+Code	Match	Static	GitHub
Voyager	Exploration	Program	Retrieve	Self-Reflection	Website
SkillAct	Demonstration	Prompt	Plan	Static	-
ASD	Task Proposal	Policy	Plan+Chain	RL	GitHub
CASCADE	Puzzle Solving	Program	Retrieve	Self-Reflection	GitHub
SAGE	Task Solving	Program	Retrieve	RL	-
PolySkill	Self-Exploration	Program	Call+Chain	Continual Learning	-
ASI	Task Solving	Program	Call	Update	GitHub
SkillWeaver	Exploration	API	Call	Collective	GitHub
EXIF	Exploration	Trajectory	-	Iterative Feedback	-
PAE	Task Proposal	Trajectory	-	RL	GitHub
Bottom-Up Agent	Exploration	Trajectory	Plan	Update	GitHub
TAIRA	Distillation	Thought Pattern	Retrieve	Reflection	GitHub
Mem^p	Distillation	Memory	Retrieve	Update	-
CER	Distillation	Memory	Retrieve	Update	-
ReMe	Distillation	Memory	Retrieve	Reflection	GitHub
LEGOMem	Distillation	Memory	Retrieve	Collective	-

Table: Summarization of representative Agent Skills implementations.

Agent Applications with Agent Skills

Domain	Method
Household	SkillAct, Mem^p
Robotics	ASD
Science	CASCADE
Coding	OpenAI Codex, Cursor, Claude Code, GitHub Copilot, VS Code, SAGE, ReMe
Web	PolySkill, ASI, SkillWeaver, EXIF, PAE, CER
Game	EXIF, Bottom-Up Agent, Voyager
RecSys	TAIRA
Travel	Mem^p
Office	LEGOMem

Table: Agent Applications with Agent Skills.

Citation


@article{wu2026agent,
  title={Agent Skills from the Perspective of Procedural Memory: A Survey},
  author={Wu, Yaxiong and Zhang, Yongyue},
  journal={Authorea Preprints},
  year={2026},
  publisher={Authorea}
}