Hugging Face AI Updates: May 26, 2026

1. Hugging Face publishes a working glossary for agent terminology that defines Agent = Model + Harness and splits scaffold from harness

Hugging Face. Hugging Face published a community glossary that tries to settle the post-ICLR 2026 confusion around “harness” and “scaffold” by grounding both terms in how production agent systems actually work, with the central formula Agent = Model + Harness as the anchor. In the proposed vocabulary the Model is the LLM itself, the Scaffold is the behavior-defining layer (system prompts, tool descriptions, context management), and the Harness is the execution layer that calls the model, handles tool calls, manages memory, and decides when to stop, all of which are governed components rather than ad-hoc code. The post also covers training-time terms that have been used inconsistently across labs — RL environments, trainers, rollouts, and rewards — and acknowledges that no definition will be universally accepted, positioning the glossary as a shared mental model rather than a standard. The piece lands as harness engineering is becoming a distinct discipline, with practitioners observing that small scaffold choices can multiply inference costs by 10× across the same benchmark. Source