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

A practical glossary clarifying commonly confused AI agent terminology, including 'harness', 'scaffold', 'policy', 'rollout', and related concepts. It distinguishes the model (the LLM itself) from scaffolding (the behavior-defining layer: system prompt, tool descriptions, context management) and the harness (the execution loop that calls the model and handles tool calls). The formula Agent = Model + Harness is explained, along with context engineering, tool use, skills, sub-agents, and training-specific terms like RL environment, trainer, rollout, and reward. The goal is to provide a shared mental model for practitioners building, deploying, or training LLM-based agents.