Skip to content

Capability, skill, and tool

These words are often used interchangeably. In a governed system, they are different layers with different owners and controls.

Term What it is Governance question
Tool A callable function or API exposed to a model Is the input schema safe? What is the blast radius?
Capability A governed unit of action with identity, policy, and audit Who may invoke it, under which conditions?
Skill A packaged procedure or knowledge artifact that an agent can use Where did it come from, and what does it contain?
Agent An actor that plans and composes capabilities toward a goal Who is accountable for its decisions?

The distinctions matter

  • A tool without a capability wrapper is an ungoverned action.
  • A capability without a skill may be too low-level for reliable use.
  • A skill without provenance is a supply-chain risk.
  • An agent without a capability catalog is guessing at what it can do.

A layered model

Agent
  └─ Skill (procedure / knowledge)
       └─ Capability (governed action)
            └─ Tool (implementation)

The layers are not mandatory, but they clarify ownership:

  • the tool owner controls the implementation;
  • the capability owner controls policy and audit;
  • the skill author controls the procedure and its dependencies;
  • the agent owner controls the goal and the plan.

Governance implications

  • Tools need input validation, sandboxing, and rate limits.
  • Capabilities need a catalog, policy, and invocation audit.
  • Skills need signing, provenance, and versioning.
  • Agents need identity, authority, lifecycle, and evidence.

Common confusions

  • “The model can call any function” — that is a tool surface, not a capability model.
  • “The skill is just a prompt” — then it has no provenance and no enforceable behavior.
  • “The capability is just an API” — then it has no policy or audit.