Digital Event Horizon
The AI agent landscape is rapidly evolving, and understanding key terminology has become crucial for practitioners. The Hugging Face glossary provides a comprehensive guide to grounding fundamental concepts, including harness, scaffold, and agent, ensuring that discussions remain clear and concise.
The Hugging Face glossary defines three fundamental concepts: harness, scaffold, and agent. A model is the core component of an AI agent, but lacks memory and cannot execute actions independently. The harness encapsulates the execution process, handling tool calls and determining when to stop. The scaffold shapes how the model sees the world and acts within it, including system prompts, context management, and tool descriptions. Context engineering designs what goes into the agent's context window, including conversation history and retrieved knowledge. A policy defines the probability of taking each possible action in a given situation, shaped by the surrounding scaffolding and harness. The glossary explores tool use, including how agents reach outside themselves using APIs, code interpreters, and databases. Skills are reusable, structured packages of knowledge that enable multi-step tasks and are portable across agents. Sub-agents are called by another agent to handle a specific subtask, with their own model and scaffold.
In the rapidly evolving field of Artificial Intelligence, the terminology surrounding AI agents has become increasingly complex and nuanced. As new frameworks, tools, and technologies emerge, it's becoming increasingly difficult for practitioners to keep up with the latest developments. To address this issue, Hugging Face has released a comprehensive glossary aimed at grounding key terms in the AI agent landscape.
At the heart of the glossary lies a critical distinction between three fundamental concepts: harness, scaffold, and agent. A model, which is essentially an LLM (Large Language Model), serves as the core component of an AI agent. However, on its own, it lacks memory and cannot execute actions independently. This is where the harness comes in – a layer that encapsulates the execution process, calling the model, handling tool calls, and determining when to stop.
The scaffold, which includes system prompts, tool descriptions, and context management, forms the scaffolding around the model. It shapes how the model sees the world and acts within it, whether during training or at inference. This distinction between harness and scaffold is crucial for understanding the different dynamics of an agent, particularly in a training pipeline.
Another critical concept discussed in the glossary is context engineering – designing what goes into the agent's context window. This includes system prompts, tool descriptions, conversation history, and retrieved knowledge. As the model runs, previous turns shape what goes into future calls, and the harness actively manages this throughout the run. Context engineering applies at both training and inference stages.
The glossary also delves into policy – a fundamental aspect of an agent's behavior. A policy defines the probability of taking each possible action in a given situation. In LLM systems, part of that policy is learned in the model weights, but it also depends on the surrounding scaffolding and harness. The same model can behave very differently depending on its prompts, tools, memory, and execution loop.
The glossary continues by exploring tool use – how agents reach outside themselves using APIs, code interpreters, databases, web search, and file systems. Modern inference APIs surface this as a first-class object, with the harness receiving the call directly and routing it to the right function.
Reusable, structured packages of knowledge called skills enable multi-step tasks. They are portable across agents and loaded on demand. The line between tool, skill, and sub-agent shifts across frameworks.
Finally, the glossary touches upon sub-agents – an agent called by another agent to handle a specific subtask. It has its own model and scaffold, reasons independently, and returns a result. This separates a sub-agent from a tool (a function call) or a skill (packaged knowledge).
The comprehensive glossary is designed to provide a practical mental model that makes discussions easier to follow. It's not meant to be a exhaustive dictionary of every term in the field but rather focuses on key concepts that are often mixed up, reused in different ways, or assumed to be obvious when they are not.
By clarifying these fundamental terms, the glossary aims to facilitate better communication and understanding among practitioners working with AI agents. It's an essential resource for those looking to stay up-to-date with the latest developments in the field.
Related Information:
https://www.digitaleventhorizon.com/articles/The-Evolution-of-AI-Agent-Terminology-A-Guide-for-Practitioners-deh.shtml
https://huggingface.co/blog/agent-glossary
https://www.mindstudio.ai/blog/agent-harness-scaffolding-matters-more-than-model
https://github.com/ai-boost/awesome-harness-engineering
Published: Mon May 25 11:17:31 2026 by llama3.2 3B Q4_K_M
