Getting AI Agent Terms Right: Harness, Scaffold, and Skills

This article tackles the messy, inconsistent vocabulary around AI agents and proposes clear definitions. Its central formula is Agent = Model + Harness. The model on its own (Claude, GPT, Qwen, or any LLM) only maps text input to text output, with no memory and no loop; what turns it into an agent is the structure wrapped around it. The author splits that structure into two layers. Scaffolding is the instruction layer: system prompts, tool descriptions, response parsing, and context management that shape how the model perceives and acts. The harness is the execution layer: it calls the model, handles tool calls, and decides when to stop. In short, scaffolding instructs while the harness executes. The piece also disambiguates terms that often get conflated. Tool Use is the model expressing intent in a structured format that the harness routes to real execution. Skills are reusable, portable knowledge packages for multi-step tasks, a level above single tools. Sub-agents are agents invoked by another agent (the orchestrator), each with its own model and reasoning. It further covers context engineering, policy, and training-side terms like rollout, reward, and rubrics. This vocabulary maps cleanly onto how MCP Nexus is layered: a Skill is a portable multi-step knowledge package, the Runner acts as the harness driving the loop, and multi-agent orchestration mirrors the orchestrator-plus-sub-agents pattern. Keeping these boundaries straight makes it easier to separate instruction from execution when building skills and workflows. 来源/Source: Hugging Face https://huggingface.co/blog/agent-glossary