Agent Systems
Agent systems are computer programs with artificial intelligence that don't just answer a goal but work through it in their own steps. To do this, they may use tools, such as a web search, a calculator, or a booking form.
A normal chat program with artificial intelligence does exactly one thing: you write something, it writes back. After that, the process is over. An agent system goes a step further. It receives a goal instead of a question, for example: “Find the cheapest train to Hamburg on Friday and add it to my calendar.” The program then works through several intermediate steps on its own, checks the result, and corrects itself if necessary. In doing so, it may also operate real tools, such as calling up a search engine or saving an appointment.
From answer-giver to task-taker
The difference sounds small, but it changes the role of the software. Until now, the AI was an advisor: it told you how you might do something. An agent system does it itself. This shifts work from the human to the machine, and not just thinking work, but also the execution.
For companies, this is the reason for the current excitement. Many office tasks consist of many small steps: extracting data from an email, transferring it into a spreadsheet, asking a follow-up question, submitting a form. Agent systems are meant to take over exactly these kinds of chains. Providers such as Microsoft, Google, OpenAI, and Salesforce have therefore been selling products they call “agents” since 2024.
But there is a downside. Anyone who allows a program to act also gives it the ability to cause harm. A misunderstood instruction then doesn’t lead to a wrong answer, but to a wrong order or a deleted file. That’s why developers build in safeguards, such as a confirmation prompt before every step that costs money.
The loop of thinking, acting, and observing
At the core of an agent system is usually a language model, that is, an AI that generates text. What’s new is what is built around this model. The model runs in a loop: it considers which step makes sense next. Then it chooses a tool and issues a command. The result of this command is fed back to it as new text, and the loop starts over.
Tools here are very concrete functions: a web search, a calculator, access to a database, a program that sends emails. The technical term for this connection is tool use. So that the model knows what it is allowed to use, it is given a short description of each tool.
Two further building blocks are common. First, a memory in which intermediate results and previous steps are stored so that the agent keeps track. Second, the division across multiple agents: one plans, one researches, one checks the result. This is called a multi-agent system. It is similar to a project team in which no one does everything alone. The price for this is more computing time and more places where an error can arise.
Where you already encounter agents today
Development is furthest along in programming. Tools such as GitHub Copilot, Claude Code, or Cursor are given a description of a task, then independently change several files, run tests, and make corrections if the tests fail. Many developers now work with them daily. This is also a good testing ground because the computer itself can check whether the result works.
In everyday life, you mainly encounter agent systems in customer service and in browsers. A support chat that doesn’t just reply but actually cancels your order is an agent. So is a browser mode that compares product prices across multiple sites and fills the shopping cart.
In the news, though, you should read the word with caution. “Agent” is currently a sales term, and some products bearing this label are just ordinary chatbots. A useful test is the question of whether the system plans its own steps and operates real tools. If it only answers, it is not an agent.