Hugging Face defines harness vs scaffold for AI agents
hugging face clarified terminology for autonomous agents, distinguishing the 'harness' execution layer from the 'scaffold' behavioral layer.
hugging face published a glossary to standardize terminology for ai agents, specifically clarifying the confusion between 'harness' and 'scaffold'. a scaffold represents the behavior-defining layer like system prompts and context, while the harness is the execution layer that handles runtime and tool calls. the article notes that while some products like claude code combine these layers, distinguishing them is crucial for engineering effective training and evaluation pipelines.