Scaffold
A scaffold is the program built around an AI model that steers its work: it breaks tasks down into steps, gives the model tools, and checks intermediate results. With a good scaffold, the same model can achieve significantly more than without one.
An AI language model can really only do one thing: continue text. You feed something in, it produces something out, and then the process is over. For many tasks that isn’t enough, because they require multiple steps. A scaffold is the program built around this model to guide it through such steps. The English word means a construction framework, and the image fits well: the model is the building, the scaffold is the construction around it that makes the work possible in the first place. It calls the model multiple times, gives it aids such as a search function, and decides when a task is finished.
Why the same model can suddenly do more
In reports on AI progress, the model is usually the focus. But a large part of the visible improvements doesn’t actually come from the model itself. It comes from the software built around it. A model that has to solve a task correctly on the very first attempt often fails. But if it’s allowed to tackle the same problem in ten small steps, test its answer, and correct mistakes, the results are noticeably better.
This has concrete economic consequences. Anyone who can’t train their own model — because that costs hundreds of millions of euros — can still build a very good scaffold. Many AI startups sell exactly that: not the model, but the control layer wrapped around it. They rent the model from a large provider and earn money on the layer on top.
Conversely, this causes disputes when companies publish benchmark figures. A model might achieve 40 percent solved tasks on a coding test on its own, and 70 percent with an elaborate scaffold. Both numbers are correct, but they measure different things. Anyone reading such comparisons should therefore ask how much supporting construction was counted in.
The components of a scaffold
The first component is the loop. The scaffold doesn’t call the model just once, but again and again. After each call, it looks at what came out and decides on the next step. This loop runs until a goal is reached or a predefined limit on the number of attempts is exceeded.
The second component is the tools. A model cannot search the internet itself, cannot open a file, and cannot run a program. The scaffold provides such functions and handles the mediation. If the model asks for a search in its response, the scaffold carries it out and feeds the result back into the next call.
The third component is control. The scaffold decides what the model gets to see in the first place, since it cannot remember an unlimited amount. It also checks intermediate results, for example by test-running written program code. If the test fails, the error message goes back to the model, and the attempt starts over. It is precisely this checking loop that distinguishes a scaffold from a merely cleverly worded prompt.
Scaffolds in products and headlines
Scaffolds are most clearly visible in coding assistants. Tools that autonomously find and fix bugs in a software project consist for the most part of scaffold code. They read files, change lines, run tests, and repeat this until everything works. The language model inside is interchangeable.
The so-called deep research features of major chat services are also scaffolds. They break a question down into sub-questions, search many web pages, and in the end write a report with sources. If such a service works for several minutes instead of seconds, that’s a clear sign of a scaffold running in the background.
A related term one often encounters is agent. This usually refers to exactly the combination of model and scaffold. A common misconception is that the model is the agent. Without the control layer around it, it can neither act nor verify anything.