Agentic Loop
An agentic loop is a repeating process in which an AI program breaks a goal down into steps, executes a step, checks the result, and derives the next step from it. Instead of a single response, this creates a chain of actions that only ends once the task is completed or aborted.
A normal AI chat program works in a single pass: you ask a question, it writes an answer, done. An agentic loop works differently. Here, the program receives a goal instead of a question, for example: “Find out which laptop under 900 euros has the best battery life.” The program then thinks up a first step, carries it out, and looks at the result. From this result it derives the next step, and the whole process repeats. “Loop” describes exactly this repetition. Only once the goal is reached does the loop stop.
From answer-giver to task-taker
The difference sounds technical, but it changes what AI can even be used for. A program that only answers once cannot notice its own mistakes. A program running in a loop can: it sees that a search turned up nothing and searches again with different words. This makes it possible to tackle tasks that consist of many small steps whose course isn’t known in advance.
This is exactly what the term “AI agent” is aiming at, a term that has appeared in almost every product announcement since 2024. At its core, an agent is nothing more than a language model, i.e. a text-prediction system, that has been wired into an agentic loop and given permission to use tools. For companies, this is the point at which AI turns from a writing aid into something that takes over work steps itself. That’s why so much money and so much expectation is riding on this concept.
The flip side is risk. A model that only outputs text can, at worst, write nonsense. A model that, within a loop, is allowed to delete files, send emails, or trigger orders can cause real damage. Errors also compound: if each step is correct 95 percent of the time, a chain of twenty steps is already predominantly wrong.
Think, act, observe
One pass through the loop usually consists of three parts. First, the model reasons in text form about what to do next. Then it chooses an action, such as a web search, running a piece of code, or opening a file. Afterward, it receives the result of that action written back into its text. This result is called an observation. This starts the next pass.
The model itself does not carry out the actions. The model only writes down which tool it wants to use and with what inputs. A surrounding program reads this request, actually calls the search engine or the database, and hands the answer back. The model is thus more like the head, the surrounding program the hands. This separation makes it possible to set boundaries for the agent — for instance: reading yes, deleting no.
So that the loop doesn’t run forever, there are termination conditions. Common ones include a maximum number of passes, a cost limit, or a check-in with a human before sensitive steps. This is necessary because agents can get stuck: they try the same failed step over and over again. Another problem is the limited working memory for text. After many rounds, the history becomes so long that the agent forgets the beginning of its task.
Agents in products and headlines
Agentic loops are most visible in programming. Tools like Claude Code, GitHub Copilot Workspace, or Cursor are given a task such as “fix this bug,” read through the code, change files, run tests, and work through the error messages. The fact that tests exist is the decisive advantage here: the agent gets clear feedback after every attempt as to whether it worked.
There are also agents that work in the browser and fill out forms, research agents that trawl through dozens of sources for a report, and support systems that look up customer data in the background. In companies, such loops often run invisibly within workflows, for instance when checking invoices.
In news reports, you should read the word “agent” critically. Often what’s behind it is a short loop with two or three tools, not an independently planning piece of software. A common misconception is also that an agentic loop is a distinct kind of AI model. It is not a model, but a construction built around one — the very same language model can answer once in a chat window or work for hours within a loop.