Back to Articles Harness, Scaffold, and the AI Agent Terms Worth Getting Right

On May 25 2026 the AI Agent community published a glossary that clarifies the often‑confused terms harness, scaffold, and sub‑agent in LLM‑based agent systems. The guide explains that a harness is the execution layer that calls the model, handles tool calls, and decides when to stop, while a scaffold is the system prompt, tool descriptions, and context‑management logic that shapes the model’s behaviour. An agent is defined as Model + Harness + Scaffold, and the document stresses that harness engineering involves stop criteria, error handling, and guardrails. Sub‑agents are separate agents invoked by a main agent to perform specialised sub‑tasks, and orchestrators manage multiple agents as units. The glossary also covers training pipelines, noting that the harness runs many episodes in parallel and feeds results back to the trainer for weight updates. The article highlights that products like Claude Code, Codex, and Antigravity CLI use the same terminology but differ in their harness implementations. By mapping existing architectures to these definitions, developers can better reason about component responsibilities and improve modularity. The guide serves as a practical mental model for both new and experienced practitioners working with agentic LLMs.