Loop Engineering

An AI chat program can answer a question in a single pass. For difficult tasks, that’s often not enough. Then the program works in rounds: it takes a step, looks at the result, and decides what happens next. This repetition is called a loop. Loop Engineering is the work of designing exactly how this loop is built. So it’s not about improving the language program itself, but about the processes surrounding it.

Why progress has shifted from the model to the loop

For a long time, the rule was: a better AI model delivers better results. That still holds true, but the gains from simply making models bigger have become expensive. At the same time, something else became apparent. The same model delivers very different quality depending on how it is allowed to work. Someone who lets it answer once gets a different result than someone who lets it check and correct itself three times.

That’s why, in many AI products today, more work goes into the loop than into the model. Companies buy the model and build the process themselves. That process is their actual product. A good example are coding assistants: they write code, run it, read the error message, and make corrections. The core of this tool is not its ability to produce text, but the cycle of trial and feedback.

The cost side also matters. Every pass through the loop costs computing time and therefore money. A loop that runs twenty times is roughly twenty times as expensive as a single answer. Loop Engineering therefore also determines whether a service can be operated economically.

The building blocks of a loop: action, feedback, termination

Every round consists of three parts. First, the model decides what it wants to do, such as starting a search or opening a file. Then this action is actually carried out, usually by a connected program. Finally, the result comes back to the model, which then plans the next round. This return flow is called feedback.

The hardest question is termination. A loop without a clear stopping rule can run forever. Typical rules include a maximum number of rounds, a time limit, or a check step that reports that the task is complete. Some systems have a second AI instance judge whether the result is sufficient. Stopping too early delivers half-finished work; stopping too late burns money.

A second core problem is memory. The model can only take in a limited amount of text at once. After many rounds, the entire history no longer fits. So one has to select, summarize, or offload intermediate results. A common misconception is that more rounds automatically produce better results. Often the opposite happens: the system loses track and repeats the same mistake over and over.

Loop Engineering in products and job postings

The loop is most clearly visible in so-called AI agents. These are systems that handle a task independently across multiple steps. When a tool reports that it solved the task in eighteen steps, you’re watching the loop at work. Research assistants that open and compare several websites on their own also work this way.

In the news, the term usually comes up in connection with cost and reliability. Reports about agents getting stuck in endless loops or generating surprisingly high bills describe failures in Loop Engineering. The topic also appears in job postings, often under titles like AI Engineer or Agent Engineer.

The term should be distinguished from Prompt Engineering, the skillful phrasing of individual instructions. Prompt Engineering concerns a single round, while Loop Engineering concerns the entire process across many rounds. In practice, the two go together, since every round needs its own instruction.