AI Agent
An AI agent is a computer program that independently breaks down a goal into multiple steps and also carries out those steps. So it doesn't just respond — it uses tools like search engines, calendars, or other programs while doing so.
Most AI programs people know from everyday life do exactly one thing: you ask a question, they write an answer. An AI agent goes one step further. It receives a goal instead of a question, for example: “Find a cheap train connection to Hamburg and add it to my calendar.” It then figures out for itself which intermediate steps are needed. Afterward, it carries out these steps by operating other programs — a search page, a calendar app, a booking system. After each step, it checks the result and decides what to do next.
From advisor to employee
The difference sounds small, but it is economically enormous. A chat program delivers text that a human then has to implement. An agent also takes care of the implementation. This shifts what companies pay for: no longer for suggestions, but for completed tasks.
That is why agents are considered the next big business model in the industry. Software companies have so far sold programs that humans operate. An agent could operate these programs itself. This explains why providers like OpenAI, Google, Microsoft, and Anthropic have built almost all their product announcements around this topic since 2024.
At the same time, a new risk emerges. A language model that writes nonsense is annoying. An agent that writes nonsense and then executes it can cause real damage — a wrong order, a deleted file, a sent email. Mistakes therefore no longer stay on the screen.
The loop of thinking, acting, and observing
At the core of an agent there is almost always a large language model. This is a program that has learned to predict meaningful text, and which can therefore also formulate plans. Around this model, a loop is built. The model writes down what it wants to do next. A piece of software executes this instruction and writes the result back into the text. Then it’s the model’s turn again.
The actions themselves are called tools. A tool can be a web search, a calculator, a database query, or an interface to another service. The developer determines in advance which tools are allowed. The agent then decides for itself which one to use and when.
On top of that comes a memory. Without notes, the agent would forget at each step what it had already tried. That’s why it stores intermediate results and reads them back in later. A comparison helps: a language model alone is like a person answering a question. An agent is the same person with a phone, a notepad, and the task of seeing the matter through to the end.
Agents in products and in the headlines
Agents are furthest along so far in programming. Tools like GitHub Copilot Workspace, Cursor, or Claude Code receive a bug description, search through the code themselves, change multiple files, and test the result. This works well there because a test immediately reveals whether it worked.
In everyday life, one encounters agents in customer service chats that not only respond but also create a return shipment. Browser features that independently fill out forms or compare prices also belong to this category. In companies, they take over routine tasks like invoice checking or appointment scheduling.
In the news, however, one should still look closely. The term is used generously, often also for software that merely follows a fixed list of rules. Another common misconception is considering agents reliable. In tasks with many steps, small errors add up, and success rates in tests are often significantly below those of humans, depending on the task. That’s why, in practice, a human confirmation is usually inserted before anything final happens.