Side by side
Tool & Agent Skill
A tool performs an operation; a skill packages instructions and resources for carrying out a kind of task.
A skill can use several tools. Tool availability and permissions remain separate from instructions in the skill.
Swap columns ⇄| Dimension | Tool | Agent Skill |
|---|---|---|
| Purpose | Execute an operation through a defined interface. | Teach a reusable way of approaching a task. [Anthropic][Agent Skills] |
| Role in the system | Callable capability. | Instructions and supporting resources available to an agent. [Anthropic][Agent Skills] |
| Inputs and outputs | Arguments go in; results or side effects come out. | Task-relevant instructions and resources guide the agent. [Anthropic][Agent Skills] |
| Persistence | Implementation-dependent; an operation may read or change stored state. | The skill package can be reused; run state depends on the application. [Anthropic][Agent Skills] |
| Example | Look up an order by its number. | Explain how to investigate an order problem using available tools. [Anthropic][Agent Skills] |
| Boundary caveats | Tool access does not by itself supply a complete procedure. | A skill may refer to tools but does not replace their execution layer. [Anthropic][Agent Skills] |
Evidence for this distinction
- Building effective agents ↗ (opens in new tab)Anthropic · 2024-12-19
Relevant section: What are agents?; Building blocks; Agents; Appendix 2 - Agent Skills Overview ↗ (opens in new tab)Agent Skills · Publication date unknown
Relevant section: What are Agent Skills?; How do Agent Skills work?
Last editorial review: 2026-09-13.