Workshop
First Workshop on Agent Skills
Design, Evaluation, and Optimization of Procedural Knowledge for LLM Agents
About
Agent Skills are structured packages of instructions, scripts, and references that augment LLM agents without model modification. They enable agents to acquire new capabilities — from tool use and API interactions to complex multi-step workflows — through composable, verifiable procedural knowledge.
This workshop brings together researchers and practitioners working on the design, evaluation, safety, and optimization of such skills across agentic AI systems. We address fundamental questions: How should skills be structured for reliability? How do we verify they're safe? Can agents learn to generate and improve their own skills?
Topics
- Agent Skills & Applications — Design principles, structure-efficacy relationships, composition and scaling
- Evaluation & Benchmarking — Frameworks, metrics, benchmark design for measuring skill effectiveness
- Safety & Supply Chain — Malicious skill detection, formal verification, prompt injection via skills
- Improvement & Learning — LLM-based skill generation, RL refinement, trajectory mining, continual learning