Agent Skills: Lightweight Procedural Memory for AI Agents
中文What
Agent Skills are an open-standard, portable directory-based primitive anchored by a SKILL.md file that equips general-purpose LLM agents with on-demand specialist capabilities. They act as a procedural memory layer, enabling agents to remember how to perform complex workflows step-by-step rather than just storing facts.
How
Skills employ a progressive disclosure architecture where only the metadata (YAML frontmatter containing name and description) is persistently kept in context. The SKILL.md body, executable scripts, and supplementary assets or references are loaded strictly on demand when a query triggers the skill. They can be written manually as markdown folders, auto-drafted by agents from successful run traces, and installed by saving them into local directories like .agents/skills/ or loading them programmatically via tools like SkillToolset.
Why
They prevent context rot and excessive token costs by loading instructions dynamically instead of bloating a single system prompt. They also offer a simpler, easier-to-maintain alternative to complex multi-agent architectures for sequential tasks, allowing a single agent to dynamically adapt to hundreds of different process variations.
Other
Skills do not compete with Model Context Protocol (MCP) or AGENTS.md. MCP provides connection to external tools and systems, AGENTS.md provides constant global project rules, and Skills provide on-demand execution know-how. Rigorous evaluation is critical, as a 2025 SkillsBench study revealed that 19% of tasks performed worse with a poorly designed skill due to trigger failures or low-quality guidance.