Packaging Agent Capabilities

Agents can receive capability through direct tools, procedural skills, and external protocol servers. These mechanisms solve different problems and carry different context and operational costs.

Plain tools for direct actions

A tool exposes one action through a stable input and output contract. Tools are appropriate for focused operations the agent may need during many tasks, such as reading a file, running a test, or querying one service endpoint.

Tool definitions remain in the active context, so keep the default set small. Prefer deferred discovery when a tool catalogue grows beyond what one role routinely needs.

Skills for repeatable procedures

A skill packages instructions, supporting references, and sometimes scripts for a recurring workflow. Skills are useful when success depends on a procedure rather than one API call.

Progressive disclosure lets the agent discover a skill from compact metadata, then load its full instructions only when needed. This keeps a large procedural library available without placing all of it in every prompt.

MCP for external systems

An MCP server exposes tools, resources, or prompts from an external system through a standard protocol. MCP is useful when capabilities belong to a separately operated service or must be shared across different agent clients.

Large MCP schemas can consume substantial context before any work begins. Load servers or individual definitions lazily where the client supports it, and expose only the surface required by the current role.

Choose the smallest fitting mechanism

Use a plain tool for a direct action. Use a skill for a repeatable procedure. Use MCP for a continuous external integration that benefits from a standard boundary. The mechanism should reduce ambiguity and context cost, not merely add another abstraction.

The agent now has instructions, a model, an action surface, and explicit authority boundaries. The next chapter uses that complete picture to examine common agent failure patterns.