Harness, Scaffold, and the AI Agent Terms Worth Getting Right

AI agents have a vocabulary problem, and Hugging Face is trying to fix it. The company just dropped a glossary that walks through terms like harness, scaffold, policy, and rollout — words that get thrown around constantly in agent circles but rarely get defined the same way twice. The project started, fittingly, with a confused question posted after ICLR 2026, when someone asked why "harness" and "scaffold" seemed to mean five different things depending on who was talking.

The core distinction the post lands on: a model is just an LLM that answers a prompt and stops, nothing more. Scaffolding is the layer that shapes how the model perceives its world — system prompts, tool descriptions, memory management. The harness is the execution engine that actually runs the loop, calling the model, handling tool calls, deciding when to quit. Products like Claude Code and Codex often just call the whole bundle a "harness," which is technically fine but blurs the line that matters most when you're building a training pipeline and need scaffold and harness to behave as separate, swappable pieces.

That swappability is the practical payoff. Two coding assistants running the exact same underlying model can feel wildly different because their harnesses make different choices about error handling, stopping conditions, and guardrails. Swap a stronger model into an existing harness, and the whole experience shifts again. Hugging Face frames it as three distinct things — model, harness, product — that people constantly collapse into one.

The glossary also covers the training-specific vocabulary that trips people up: environments (the stateful thing an agent acts on and gets observations back from), trainers (the system that runs episodes and updates weights, with TRL's GRPOTrainer as the example), rollouts (a full agent run, also called a trajectory), and rewards, which can be verifiable, learned, sparse, or dense. Sub-agents get a clean separation from tools and skills too — a sub-agent reasons independently and can call further sub-agents, while a tool is just a function call and a skill is a packaged bit of know-how.

None of this is presented as gospel. The authors are upfront that different frameworks use these words differently and that plenty of terms don't have settled definitions yet. The point isn't to crown one correct dictionary — it's to give people enough shared footing that a conversation about agent harnesses doesn't turn into three people talking past each other.