An LLM by itself is a stateless function that produces text and then stops. To transform a model like Claude, Qwen, or GPT into an agent, you must wrap it in two distinct layers: scaffolding and a harness.
Scaffolding defines the model's world
Scaffolding acts as the behavior-defining layer that surrounds the model. It includes the system prompt, tool descriptions, and the logic for context management—deciding what the model remembers across multiple steps. It essentially shapes how the model perceives its environment and dictates the format of its responses.
When developers talk about products like Claude Code or Codex, they are often referring to the entire integrated stack. However, in a rigorous engineering or training pipeline, scaffolding is specifically the instructions and data structures the model works from. It is the "software" that guides the model's intent.
Harness enables the execution loop
The harness is the execution layer that actually makes the agent run. While the scaffolding provides the instructions, the harness handles the heavy lifting: calling the model, intercepting tool calls, managing errors, and deciding when the agent's task is complete.
This distinction is critical for understanding why two products using the same underlying model can feel entirely different. The difference lies in the harness engineering—the specific choices made regarding guardrails, error handling, and the loop logic. If you are not the model, you are the harness. This principle extends to evaluation, where an "eval harness" runs fixed scenarios against model checkpoints to record metrics without updating weights.
The Agent equation
In the modern LLM landscape, the community has largely converged on a simple mental model: Agent = Model + Harness. The scaffolding is the content within that harness that the model consumes.
By separating these layers, engineers can reason about them independently. You can swap a more capable model into an existing harness to improve performance, or you can redesign a harness to change how an agent interacts with its environment without ever touching the underlying model weights. This modularity is what enables the rapid evolution of agentic workflows in coding, research, and automated tool use.
Source: Harness, Scaffold, and the AI Agent Terms Worth Getting Right
Domain: huggingface.co
Comments load interactively on the live page.