Hugging Face wants the agent vocabulary to settle

If you have been confused about whether Claude Code is an "agent", a "harness", or "scaffolding", you are not alone. Hugging Face's new glossary post by Sergio Paniego and Aritra Roy Gosthipaty argues the agent vocabulary has drifted faster than people can use it. At ICLR 2026 this year, the authors note, practitioners could not agree on basic definitions. The post is an attempt to fix that.

The framing they propose is clean. The model is just the LLM: text in, text out, no memory, no loop. Scaffolding is the behaviour layer wrapped around it, including the system prompt, tool descriptions, response parsing, and context management. The harness is the execution layer that calls the model, runs tool calls, and decides when to stop. An agent is the whole stack working together in an observe-decide-act loop.

The most useful part is the distinction they draw between scaffolding and harness. The same model under different scaffolding behaves like a different product, and the harness is what makes any of it actually run. Claude Code, Codex, and Cursor are all harnesses around models, but each makes different architectural choices about what the model sees and when it should stop. The post also covers tools versus skills, sub-agents, context engineering, and the training-side terms (RL environment, rollout, reward, trainer) for people building agents from scratch.

Why it matters

If your team is shipping agents or writing about them, agreeing on these four words (model, scaffolding, harness, agent) will save more time than you might guess. It also makes RFP conversations and post-mortems cleaner. The bug is usually in the harness or the scaffolding, not in the model itself.