Concept·Agents & Automation·Added 1 day ago

Scaffold

Also known as: agent scaffold, scaffolding, LLM scaffold, model scaffold

The code you write around a model to make it actually do things: routing its outputs back into inputs, giving it tools to call, managing loops. The model is the brain; the scaffold is the nervous system connecting it to the world. Often confused with harness, which is a broader concept.

A language model by itself just takes text in and returns text out. A scaffold turns that into a working agent by wiring the model into a loop: it gets a task, calls a tool (like a web search or a code runner), sees the result, thinks about the next step, calls another tool, and so on until the job is done. The scaffold manages that cycle, tracks state, formats tool outputs, and decides when to stop.

Scaffold is sometimes used loosely to mean all the code surrounding the model, and sometimes more precisely to mean just the loop-and-tool-routing layer, as opposed to the harness (which covers the full operational environment including monitoring, memory, and retrieval). The distinction isn't settled in the industry. A useful rule: if someone is describing what makes agents actually run, they're probably talking about scaffolding.

Tools like LangGraph, CrewAI, and OpenAI's Agents SDK are scaffold frameworks: they give you scaffolding primitives (loop management, tool calling, state tracking) so you don't have to write all that plumbing from scratch. When you hear 'I built my own scaffold,' it usually means someone rolled their own agentic loop rather than using a framework.

This definition is AI-generated and refreshed weekly. It may contain inaccuracies. Use your own judgment, especially for production decisions.

Related terms