Harness, Scaffold, and the AI Agent Terms Worth Getting Right

This article clarifies terminology used in the AI agents field by providing practical definitions for often-confused concepts.

• •Scaffolding is the behavior-defining layer (system prompt, tool descriptions, context management) that shapes how the model sees and acts in the world
• •Harness is the execution layer that calls the model, handles tool calls, and decides when to stop the loop
• •An agent is a model plus everything around it (harness and scaffolding) that enables it to act in a loop
• •Context engineering involves designing what the model sees at each step, including system prompt, tool descriptions, conversation history, and retrieved knowledge
• •Tool use, skills, and sub-agents are distinct concepts where tools are actions, skills bundle knowledge for multi-step tasks, and sub-agents reason independently