Papers on Agent Skills

Four-axis survey; SKILL.md spec; four-tier trust framework; 26.1% vulnerability rate

Lifecycle systematisation; quality ceiling problem; concept unification table

200K+ skills; three-layer ontology; SCECM evaluation; 40% reward improvement

Capability tree; DAG orchestration; ablation proving retrieval+orchestration both needed

84 tasks x 11 domains; modular > exhaustive; self-generated sometimes > human

Discover -> practice -> synthesise -> hone; 31.8% WebArena improvement

RL-optimised skill library; sequential rollout; 8.9% higher completion

Curriculum-based self-evolution; 23.2pp gain over UI-TARS

Exploration-first; Alice discovers, Bob learns; iterative feedback

Experience -> insight extraction; transfer learning; no parameter updates

Workflow structure as first-class optimisation target

Dynamic workflow revision during execution

Heterogeneous workflows via evolutionary algorithms

RL-trained meta-agent generates per-query workflows

Workflow construction as grouped think-act RL

First skills-as-artifacts benchmark; 84 tasks x 3 conditions

Containerised terminal agent evaluation

Production-level code issue resolution

Human-verified subset of SWE-bench

Extended software engineering tasks

Self-supervised tool use learning

Massive API mastery framework

Comprehensive tool learning survey; four stages

Automated data generation for tool calling

Tool-augmented RL; 67% AIME2024 accuracy

Entity-relation graphs from corpora; community summaries

First comprehensive GraphRAG overview

Unified GraphRAG formalization

IEEE survey on graph-based RAG

Hierarchical retrieval interfaces; agentic autonomy

Factual/experiential/working memory taxonomy; dynamics

Cognitive mechanism + memory subject perspectives

Inside-trial vs cross-trial information; memory management

Bayesian selection; meta-procedural playbooks; 78.1% avg

Systematic procedural memory analysis

Tool creation spectrum; self-evolution taxonomy

GVU operator; variance inequality for self-improvement

Metacognitive knowledge + planning + evaluation

Self-play bug injection and repair; +10.4pp SWE-bench

Methods for agent self-improvement

Collaboration taxonomy: actors, types, structures, strategies

Unified understanding of coordination across domains

Delegation, supervision, cooperation, coordination

Manager Agent for ad hoc teamwork

MCP + A2A dual protocol foundation

Prompt chaining, routing, parallelization, orchestrator-workers

Step-by-step prompt engineering

Self-generating instruction examples

Tree-based deliberation for complex reasoning

Step-by-step reasoning in LLMs

Hierarchical doc IDs; prefix-sharing

Completeness-oriented retrieval for LLMs

Memory-augmented recommendation agent

Episode pagination -> memory gisting -> lookup

Dynamic tree memory for LLMs

World model for LLM planning

Critical investigation of LLM planning

Comparison with classical planning

Task planning with tool usage

Three-paradigm review; CoLing 2025

26.1% skills contain vulnerabilities

Skill-file prompt injection risks

MCP, Agent Cards, related protocol comparison

Benchmarking vulnerability discovery and remediation

Security, privacy, ethics threat survey

HIL evaluation for MCU programming skills

MCP for IoT system integration

ReAct-style with explicit memory for web

Autonomous visual tool invocation for GUIs

Attention-based coordinate-free GUI grounding

Definitive skills survey; four axes

Lifecycle stages; quality ceiling

Tool creation spectrum; self-evolution taxonomy

Agent memory taxonomy and dynamics

Cognitive + subject perspectives on memory