AI Agent

Most well-known AI programs only respond. You ask a question, the program delivers text, and then the matter is done. An AI agent goes a step further. It receives a goal instead of a single question and then works independently toward it. To do so, it is allowed to do things: search the internet, launch a program, save a file, send an email. Along the way, it checks how far it has gotten and decides for itself on the next step. The difference, then, lies not in intelligence but in freedom of action.

From answering to getting things done

A pure question-and-answer system merely shifts the work. It explains to you how to book a trip, but you have to book it yourself. An agent is meant to take over this last part. This is exactly where many companies see the economic value: it's not the advice that's expensive, but the hours someone spends on implementation.

That's why, since around 2024, a great deal of money has been flowing into this area. Software vendors sell agents that handle customer inquiries, check invoices, or write program code. Increasingly, billing is done per completed task rather than per user and month. This is changing the business model of the entire industry, which is why the term keeps popping up in financial news as well.

The flip side is risk. A program that only outputs text can, at worst, assert nonsense. A program with access to accounts, calendars, or servers can cause real damage. An agent that deletes the wrong file has truly deleted it. That's why developers build in safeguards, such as a confirmation prompt before any action that cannot be undone.

The loop of thinking, acting, and observing

At the core of an agent usually sits a language model, i.e., an AI system that continues text and, in doing so, understands instructions. This model receives a description of the goal and a list of permitted tools. A tool is a clearly described function, such as a web search or a calculator. The model then writes not flowing text but an instruction like: Tool web search, search term train connection Hamburg Berlin.

A surrounding program executes this instruction and passes the result back to the model. Then the process starts over from the beginning: the model looks at the result, thinks further, and chooses the next step. This loop of thinking, acting, and observing runs until the goal is reached or an upper limit on steps is exceeded. It is often called an agent loop.

On top of this comes memory, since the model itself remembers nothing between two calls. The agent therefore keeps a log of all steps taken so far and feeds it back in every time. For long tasks, this log is summarized, because otherwise there isn't enough space within the model. This is precisely where agents often fail: after many steps they lose the thread or repeat themselves endlessly.

Agents in products and headlines

Agents are furthest along in programming. Tools like GitHub Copilot, Claude Code, or Cursor receive a task such as "fix this bug," search through the files themselves, change code, and run tests. This works well there because success is measurable: the test either passes or it doesn't. The agent thus gets honest feedback immediately.

In everyday life, you encounter agents in browsers that click through websites themselves, and in customer portals that handle a return from start to finish. The term computer use also comes up often: it means that the AI sees a screen and controls the mouse and keyboard. In the news, you can read figures on how many tasks an agent completes without human intervention.

A common misconception is that agent is simply a new word for chatbot. The difference is verifiable: a chatbot delivers text, an agent changes something outside the conversation. Equally misleading is the notion of complete autonomy. In practice, almost all agents operate within tight limits, with fixed tools, budgets, and a human who confirms important steps.