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

The rapid advancement of AI agents has created a vocabulary that often evolves faster than shared understanding, leading to confusion around terms like harness and scaffold. According to Hugging Face, these concepts frequently blur or get reused inconsistently across different frameworks. To address this, a new glossary aims to establish practical mental models for key terms without enforcing a single universal definition. In this framework, the model is strictly the underlying LLM, which possesses no memory between calls and requires external elements to function as an agent. Scaffolding represents the behavior-defining layer around the model, including system prompts, tool descriptions, formatting parsers, and context management. Meanwhile, the harness acts as the execution layer that drives the model through its operational loop, handling tool calls and determining stopping conditions. Together, the model, scaffolding, and harness comprise an AI agent capable of taking in information, executing loops, and acting on results. The glossary also covers related concepts like harness engineering, context engineering, memory management, policies, tools, skills, and sub-agents, noting that differences between frameworks often stem from how these components interact. By distinguishing between the model, the harness, and the product itself, practitioners can better reason about development pipelines, training processes, and inference deployment.