Agentic Models
Agentic models are AI systems that don't just answer a task but independently break it down into steps and work through those steps using tools such as web search or program code. In other words, they act instead of merely responding.
Most AI programs known from everyday life work like a question-and-answer game. You type something in, the program returns a text, and that’s it. Agentic models go a step further: they are given a goal and work toward it independently. To do this, they break the goal down into several sub-steps and complete them one after another. Along the way, they are allowed to use tools, such as a search engine, a calculator, or a booking form. In between, they check whether the result fits and correct themselves if necessary. So the difference isn’t intelligence, but autonomy.
From advisor to employee
A pure answering system can explain to you how to book a flight. An agentic model is supposed to actually book the flight. That’s exactly where the economic appeal lies. Tasks that a human currently handles at a computer could partly shift over to software. That’s why companies like OpenAI, Google, and Anthropic have been investing heavily in this direction since 2024.
For companies, this shifts the billing logic. Previously, you paid for a tool that made the work easier. With agents, it’s about the work itself, meaning hours instead of licenses. Analysts see this as the reason why the stock market talks so much about this topic.
At the same time, the risk grows. Anyone who lets a system act must expect that it will sometimes act wrongly. A text error is annoying, a wrong bank transfer is expensive. That’s why many agents today only run with human approval at critical points.
Plan, act, check
At its core, there is almost always a language model, meaning an AI system that has learned to continue meaningful texts. The trick is to run this model in a loop. The model first writes down what it wants to do next. Then the surrounding program executes this action. The result is fed back to the model, and the round starts over.
The actions are so-called tool calls. The model might write, for example: search the web for the current exchange rate. The program calls the search engine and returns the result. This way, an agent can also obtain information that didn’t even exist during training. Other typical tools are a file system, a browser, or the execution of program code.
You can picture this like an intern with a task list. He checks off point after point, is allowed to look things up in between, and reports back when he gets stuck. The big weakness is error propagation: a wrong step in round three ruins all the following rounds. With twenty sub-steps, even a small error rate per step is enough to make the overall result useless.
Agents in the browser and in programming
Agentic models are furthest along in programming. Tools like GitHub Copilot Workspace, Claude Code, or Cursor receive a bug description and then independently change several files in the project. Afterward, they run the tests and check whether everything works. Programming is well-suited for this because the result can be verified by machine.
The second major area is browser agents. They control a website the way a human would with a mouse and keyboard, meaning clicking and typing. This allows forms to be filled out, prices to be compared, or appointments to be entered. In practice, though, such agents still often fail at login windows and cookie banners.
In the news, you’ll usually encounter the term as a promise for the future. In that case, pay attention to two questions: how many steps can the agent manage without help, and what is it allowed to do without asking first? A system that only makes suggestions is something completely different from one with access to accounts. It’s exactly at this boundary that it’s decided whether agentic models are useful or dangerous.