Hugging Face Publishes AI Agent Glossary to Standardize "Harness" and "Scaffold" Terminology

Summary

Hugging Face published an agent terminology glossary on May 25, authored by Sergio Paniego and Aritra Roy Gosthipaty, motivated by terminology confusion that surfaced at ICLR 2026. The glossary codifies the practitioner consensus that has been emerging through 2026: an agent is the combination of a model (the LLM), a harness (the execution layer that calls the model, handles tools, and decides when to stop), and a scaffold (the behavior-defining layer — system prompts, tool descriptions, respons