Agent Scaffold

An AI language model can, by itself, do only one thing: generate the likely fitting next piece of text from a given text. It cannot open a file, cannot start a search, and cannot remember anything beyond the current conversation. An agent scaffold is the program built around this model so that it nevertheless becomes a helper that works independently. It calls the model again and again, carries out what the model proposes, and feeds the result back. The English technical term for this is scaffolding, as in a construction scaffold. The image fits well: the scaffold itself doesn’t think, but without it you couldn’t reach the higher floors of the building.

Why the same model can suddenly do more inside a good scaffold

Two companies can buy the same language model and yet offer very different products. The difference often lies not in the model but in the scaffold built around it. This was clearly observable with coding assistants: the same model version solved significantly more test tasks than before once placed in an improved scaffold. All that had changed was how often the model was allowed to test its own code and how error messages were fed back to it.

For companies, this is important news. Training a better model costs enormous sums and takes months. Building a better scaffold is ordinary software work and can be finished in weeks. That is why a large part of the competition in the AI industry is currently happening not around the models themselves, but around the software built around them.

The scaffold is also the place where safety is created. It decides whether an agent may actually delete a file or trigger a payment. You can ask a model to be careful, but that cannot be guaranteed. A block built into the scaffold, on the other hand, holds firm, because it is ordinary program code.

The loop of thinking, acting, and observing

At its core, almost every scaffold operates as a loop. It gives the model the task and a list of permitted tools, such as a web search or a calculator. The model responds with a proposal for which tool it wants to use and with what inputs. The scaffold carries out this step and sends the result back as new text. Then the next round begins, until the task is complete.

On top of that comes memory management. A model can only read a limited amount of text per request; this is called the context window. In a task with fifty steps, at some point not everything fits anymore. The scaffold therefore summarizes older steps, stores intermediate results in files, and retrieves only what is needed.

A common misconception is that the scaffold is merely a nice frame for input instructions. In fact, it contains the trickiest decisions. How many times may an agent repeat a failed step before giving up? Without such abort rules, agents spin in circles for hours and cause real costs, since every model call costs money.

From coding assistant to booking helper

Scaffolds are best known from coding tools. When an assistant independently reads files in a project, writes a change, and then runs the tests, exactly this kind of loop is behind it. Research agents work similarly, calling up dozens of web pages for a report and merging the results.

In companies, scaffolds are increasingly appearing in customer service and administration. An agent reads a complaint, looks up the order in the database, and creates a return slip. The last step, the refund, is often deliberately blocked and requires approval from a human. This boundary is drawn by the scaffold, not by the model.

In the news, you mainly encounter this term in benchmark reports, that is, reports about performance tests. They state that a model solved a certain percentage of tasks. Such figures are only comparable if the scaffold is also the same. Anyone who leaves this out is, in truth, comparing two different systems.