Agentic AI
Agentic AI refers to AI systems that don't just answer a task but break it down into steps and work through them independently. They plan, use tools like search engines or programs, and check their intermediate results until the goal is reached.
Most AI programs people are familiar with work according to a simple pattern. You ask a question, the program answers, and then the process is over. Agentic AI goes a step further. Here the system is given a goal instead of a question, for example: “Find the three cheapest train connections to Hamburg and add them to my calendar.” The system then figures out for itself which sub-steps are necessary and carries them out one after another. In doing so, it may also operate other programs, such as a search engine or a calendar app.
From answer-giver to task-taker
The difference sounds small, but economically it is large. A chatbot saves a person a few minutes of thinking. A system that completes an entire task from start to finish saves an entire work step altogether. This is precisely why software companies are currently investing very large sums of money in this direction.
Typical use cases are tasks made up of many small, tedious manual steps. Finding invoices in emails and filing them. Checking customer inquiries and routing them to the right department. Writing code, testing it, and fixing the bug that the test reported. A human can handle such chains without any problem, but they cost time.
At the same time, a new risk emerges. If a chatbot writes nonsense, you read it and ignore it. If an agent does something nonsensical, it may already have triggered an order or deleted a file. So errors have consequences in the real world, not just on the screen.
Planning, using tools, revising
At its core there is almost always a language model, that is, an AI system that has learned from vast amounts of text to generate meaningful text. This model is built into a loop. It first considers what the next step should be. Then it carries out that step, looks at the result, and decides again. This loop keeps running until the goal is reached or a set limit kicks in.
What matters is the tools the system is allowed to access. It is given a list of functions, such as “search the internet,” “read this file,” or “send an email.” The model then doesn’t simply write text but generates a call to one of these functions. A program wrapped around it actually executes the call and passes the result back.
A common misconception is that Agentic AI is a completely new kind of model. That’s not true. It is usually the same language model as in the chatbot, just embedded in an environment with memory, tools, and a repetition loop. Newer models, however, are specifically trained to formulate such tool calls cleanly. Because every loop iteration costs computing time, agents are also considerably more expensive to run than a single answer.
Agents in products and headlines
Agentic systems are furthest along in programming. Tools like GitHub Copilot or Claude Code are given a task description and then independently work on multiple files of a project. They run tests, read the error messages, and try again. There’s a simple reason why software of all things is leading the way: a test immediately and unambiguously tells you whether the result works.
In the browser, there are now agents that click on web pages, fill out forms, and book trips. So far these offerings work less reliably, because websites keep changing constantly. Many providers therefore ask for confirmation again before critical steps such as a payment.
The term currently appears very frequently in business news, often as a selling point. Not every product called an “agent” actually works independently. A useful test question is: Does the system make several decisions in sequence without a human intervening in between? Only then is the term justified.