Agent, Sub-Agent, Skill, or Tool? A Practitioner's Guide to Extending Agentic AI Systems

When designing a new AI capability, practitioners face a fundamental architecture decision: should it be implemented as a tool, a skill, a sub-agent, or a full agent? This seemingly simple question has profound implications for system behavior, debuggability, and maintainability. Yet these terms are often used interchangeably in industry discourse, obscuring the most critical design axis: control over execution. This paper provides clear definitions for tools, skills, sub-agents, and agents based on their control characteristics rather than their perceived intelligence. We introduce the emerging "Skills" abstraction-lightweight, portable modules of procedural knowledge that achieve significant context efficiency gains through progressive disclosure. We present canonical orchestration patterns with practical guidance on when to employ each. Through analysis of prominent frameworks including LangChain, AutoGPT, HuggingGPT, and Voyager, we demonstrate how this taxonomy maps to real implementations. We present detailed case studies from healthcare, gaming, and enterprise domains, each with architectural diagrams illustrating the orchestration decisions. We conclude with a decision framework to help practitioners select the appropriate construct, arguing that autonomy should be earned through necessity rather than assumed by default.