scaffolding

advancedwidespread

Scaffolding is the software framework around the language model that provides it with tools, memory, and a workflow to perform tasks independently.

In plain terms

A language model by itself only receives text and returns text. It cannot open files, search the web, or remember what it did yesterday. Scaffolding is the program layer built around it that makes all this possible.

Think of an excellent chef standing in an empty room. They have the knowledge, but without a kitchen, tools, ingredients, and order slips, there will be no dinner. Scaffolding is the kitchen: it does not cook, but without it, the chef's knowledge does not become a meal.

This is why the same model only chats in one product, while in another it works as an agent on a task for hours. The difference is often not in the model, but in the scaffolding built around it.

How it works

The framework compiles what the model sees: the task, the rules, the description of available tools, and the results of previous steps.
The model indicates what it wants to do, such as searching, reading a file, or running a program.
The framework actually executes this request and then returns the result to the model.
This loop repeats until the task is completed or the framework stops the work due to a step, time, or cost limit.
Meanwhile, the framework manages memory, compresses long histories, and, if well configured, asks for your permission before risky steps.

An everyday example

You ask your AI assistant to compare three insurance quotes from attached PDFs and create a table. Opening the files, passing their text to the model, and saving the table are handled not by the model but by the scaffolding built around it. The model interprets the quotes, formulates the table content, and decides what the next step should be.

Who is affected

developerAnyone building an agent spends most of their work designing the scaffolding: tools, memory, error handling, and shutdown conditions. → Code

entrepreneur, executiveWhen choosing an AI product, not only the model matters, but also the framework in which it runs, as this determines its capabilities and reliability.

employeeIf you understand that your assistant's capabilities depend on the framework, you will know more precisely what to entrust to it and what not to. → Work & HR

A common misconception

MythIf one AI tool is more skilled than another, it definitely contains a smarter model.

In realitySometimes the same model powers both, just surrounded by different scaffolding. Better tools, better memory management and better instructions can make a dramatic difference.

Limits and risks

Scaffolding does not make the model smarter: if the model makes a bad decision, the framework at most detects or limits the error.
The more tools and permissions the framework grants, the greater the damage a mistake or a malicious, hidden instruction can cause.
Multi-step work is slower and more expensive than a simple question-and-answer, because each round requires a new model call.
Comparative tests of models can be misleading if they were not measured in the same framework.

Similar, but not the same

language model – The model is the 'brain' that generates text, while scaffolding is the traditional code written around it, providing hands, memory, and workflow.
fine-tuning – Fine-tuning changes the model itself through additional training, while scaffolding does not touch the model, only shapes its environment.
educational scaffolding – In pedagogy, it refers to the teacher's gradually withdrawn assistance, while in AI it is a software framework that permanently remains around the model.
code-generating scaffolding – In web development, this term has long been used for when a framework (such as Ruby on Rails) automatically generates the skeleton of an application. This has nothing to do with the framework built around AI agents.

Related concepts

Broader

AI-agent, agentrendszer

Narrower

tool use, agentciklus, memory, system prompt

On the edge

orchestration, context engineering, prompt chain, RAG, large language model

A brief history

The word originates from construction scaffolding, and was transferred to pedagogy in 1976 by Wood, Bruner, and Ross in their study. In AI, it spread around 2023, when the first autonomous agents built on language models appeared, such as AutoGPT.

Where things stand

At the end of 2026, virtually every serious AI assistant and coding agent has mature scaffolding behind it. In the industry it is increasingly called a 'harness', and it is tracked separately when models are evaluated.

Status: 2026 November

In the news

Appears in 16 Agent Age stories; the curve shows weekly occurrence in the news magazine.

GitHub Copilot Fusion mixes models to approach frontier coding performance at half the costOct 28, 2026 · Agent Age
Copilot Studio's GitHub Copilot harness becomes the Microsoft Copilot harnessOct 28, 2026 · Code
IFS launches industrial AI products for reliable execution and queryable operational dataOct 13, 2026 · Ipar
OpenAI's Agents API offers Codex harness as a service with computer use; data stored only in USSep 30, 2026 · Code
GitHub Security Lab automates fuzzing with an AI agent frameworkSep 24, 2026 · Code
OpenAI Agents API in public beta: the Codex harness for cloud agentsSep 21, 2026 · Code
Harness report: AI speeds up development, but deployment failure rate is 69%Sep 15, 2026 · Code
OpenAI opens the Codex harness to developers as a managed Agents APISep 14, 2026 · Code

Also known as

scaffolding · scaffold · agent scaffolding · agent scaffold · harness · agent harness · scaffolding · agent framework

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.