Agent skills

Agentic AIArchitecture and orchestrationPublished By Simon Budziak

Agent skills are reusable packages of instructions, examples, scripts, and references that teach an AI agent how to perform a specific kind of work. The agent sees a short description first and loads the detailed procedure only when the current task needs it.

The format turns procedural knowledge into a maintained asset instead of repeating it inside every prompt. The Agent Skills specification describes a folder built around a SKILL.md file, while current frameworks can expose those folders through an agent harness.

How do agent skills work?

A harness advertises each skill using a compact name and description. When a task matches, the agent loads the full instructions and any referenced resources. Progressive loading keeps unrelated procedures out of the context window, which matters when a coding agent has dozens of available workflows.

How are skills different from tools?

A tool performs an action. A skill explains how and when to use tools together. An agent may follow one skill while making several tool calls, including calls exposed through MCP. Teams should review third-party skills like code because instructions and bundled scripts can both change agent behavior. Our LangChain agent skills guide shows how that separation works in practice.

Frequently asked questions

How is an agent skill different from a prompt?

A prompt gives instructions for one interaction. A skill packages a reusable procedure with supporting examples or files and can be discovered when a relevant task appears.

Are agent skills the same as MCP tools?

No. A skill teaches an agent how to approach work, while an MCP tool gives it a standardized way to call an external system or retrieve data.

Summarize this page with

See this working in a system we built