Glossary Clarifies AI Agent Terms: Harness vs. Scaffold

A new glossary from Hugging Face aims to clarify the often-confused terms 'harness' and 'scaffold' in AI agent development. Authored by contributors including @ariG23498, the glossary emerged from confusion at ICLR 2026, where practitioners could not agree on definitions. The glossary distinguishes the model (LLM like Claude, Qwen, GPT) from the harness (execution layer handling tool calls and loops) and scaffolding (behavior-defining layer: system prompt, tool descriptions, context management). It notes that products like Claude Code and Codex use 'harness' broadly to mean everything outside the model, while scaffold/harness distinction matters in training pipelines. The glossary also covers context engineering, memory, policy, tools, skills, sub-agents, and training-specific terms (environment, trainer, rollout, reward). It references frameworks like TRL's GRPOTrainer and resources like the HF Context Engineering Course. The authors thank reviewers Pedro Cuenca, Quentin Gallouédec, Shaun Smith, and Adithya S Kolavi.