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

The rapid evolution of the AI field often leads to a vocabulary that develops faster than shared understanding. Terms become blurred, are reused in different contexts, or act as shorthand for unexplained ideas. This is particularly evident in AI Agents, where concepts are frequently mixed, renamed, or disappear after a short period. This can be overwhelming for both newcomers and experienced practitioners.

This article aims to clarify frequently misunderstood AI agent terms. It focuses on concepts that are often confused, reused inconsistently, or assumed to be obvious. The goal is not to enforce a single vocabulary but to provide a practical mental model for easier discussions.

A "model" is the Large Language Model (LLM) itself, processing text input to output. On its own, it lacks memory and looping capabilities. A "scaffold" is the behavior-defining layer surrounding the model, including system prompts and tool descriptions. It shapes how the model interacts with its environment.

The "harness" is the execution layer within an agent, responsible for calling the model, handling tool calls, and deciding when to stop. While some products broadly define "harness" to encompass everything beyond the model, the distinction between scaffold and harness is crucial for separate reasoning, especially in training pipelines. "Harness engineering" focuses on designing this execution layer effectively.

An "agent" combines a model with its surrounding components that enable action beyond mere responses. It transforms raw text generation into a cyclical process of information intake, decision-making, and action. For example, in a coding agent, the system prompt and tool descriptions form the scaffold, while the loop managing model calls and tool handling is the harness.

"Context engineering" involves designing what information the agent sees at each step within its context window. This includes the system prompt, tool descriptions, conversation history, and retrieved knowledge. This is an ongoing process, not a one-time decision.