Tools and skills

Tools and skills are what an agent can do. The two words are paired and not interchangeable. A tool invokes an external or code-defined capability through an approved interface — such as a direct API, SDK, or Model Context Protocol (MCP) server. A skill runs a model-mediated transformation such as summarization, classification, extraction, or drafting.

Why the distinction matters

The distinction is operational. Tools have explicit interfaces, permissions, side effects, and error behavior; they are not automatically deterministic just because they use a protocol. Skills are model-mediated and can vary across runs. Treating both as unbounded “functions an agent can call” makes agentic systems hard to inspect and debug.

What an agent has access to

An agent in a customer-specific system design receives only the tools and skills justified by its scope. One agent may receive an approved read-only source while another may receive a separately authorized write capability; neither inherits the other's authority. MCP is one possible interface for those capabilities, not a requirement. Scoping is what makes the agent's behavior legible.

How they are versioned

Tools and skills use versioned interfaces in an approved system design. When a tool's output schema changes — a new field, a deprecated one — consumers handle the new version explicitly. Production routing does not move silently across versions.