AI Agent

An AI agent is a program that independently breaks a task down into sub-steps and carries out those steps. You don’t give it a single question, but a goal: “Find the cheapest train to Hamburg on Friday and book it.” The program then works out for itself which steps are needed for this. In doing so, it may access tools, such as a search engine, a calculator, or a booking website. After each step, it looks at the result and decides what makes sense to do next. This cycle of reasoning, acting, and checking continues until the goal is reached or the agent gives up. In German, the feminine form 'KI-Agentin' is also used – it refers to the same thing.

From answering to getting things done

Until recently, the usefulness of AI systems was mainly linguistic. They wrote texts, summarized, translated. The work that followed remained with the human: copying, pasting, pressing buttons. Agents are meant to take over exactly this part. That is the reason why large technology companies built almost all their product announcements around this word in 2024 and 2025.

Economically, the difference is significant. A company pays little for a written email. For a fully processed customer case, it pays considerably more. That’s why providers are shifting their pricing models: no longer per user and month, but per completed task. Analysts see this as the transition from a tool market to a labor market.

But this also increases the risk. A wrong sentence in a text is embarrassing. A wrong bank transfer or a deleted database is real damage. An agent that is allowed to act can not only express errors but carry them out into the world. This is exactly what research is currently working on most intensively.

The cycle of planning, tools, and control

At the core of an agent there is usually a language model. This is an AI system that has been trained to continue texts and, in doing so, is very good at understanding instructions. This model doesn’t generate flowing text for humans, but short action commands: “search for X,” “read file Y.” A surrounding program reads these commands and actually executes them.

The executable actions are called tools. A tool can be a web search, a database query, a program that runs code, or an interface to a calendar. The result of each tool call is returned to the model. This creates a loop: thought, action, observation, next thought.

To keep track of things, the agent needs a memory. In the short term, this is the conversation history so far; in the long term, a storage where intermediate results and notes are kept. For larger tasks, several agents sometimes share the work, for example one for planning and one for checking. The most important building block, however, is still often the human: at critical points, the agent asks for approval before spending money or deleting data.

Agents in products and headlines

The technology is furthest advanced in programming. Tools like GitHub Copilot Agent, Claude Code, or Cursor receive an error message, search the project themselves, change files, and run the tests. This works so well because success is measurable: tests are either green or red. So the agent notices by itself when it’s wrong.

There are also agents that click and type in the browser, enter invoices into accounting systems, or handle customer inquiries all the way to a refund. In news reports, you often come across terms like “agentic AI” or “computer use.” A common misconception is that such systems want something. They merely pursue a goal set by a human and have no interest of their own.

When reading announcements, a critical question is worth asking: how often does the agent complete the task without help? In studies, success rates for multi-step office tasks are often still well below fifty percent. Much of what is sold as an agent is, so far, a chatbot with two tools attached.