Use a tool for one callable action. Use MCP when an agent needs a standard connection to live external data or capabilities. Use a skill for a reusable workflow with instructions, references, and scripts. Use a specialized agent when independent reasoning or isolated context improves the task.
These pieces can combine: a skill tells an agent how to review incidents, MCP supplies the traces, and a subagent analyzes a bounded cohort.
Choose the smallest surface that matches the need. Unnecessary layers increase configuration, permissions, and debugging cost.