Agent Mode

Most AI chat programs work according to a simple pattern: you write something, the program replies with text. After that, the process is over. In Agent Mode, things work differently. There, the program receives a goal instead of a question, for example “find the three cheapest train connections and put them in a table.” It then breaks this goal down into intermediate steps itself and works through them one by one. In doing so, it may use tools: calling up a search engine, saving a file, performing a calculation. The human thus only gives the task and looks at the result at the end.

From advisor to executor

The difference sounds small, but it is fundamental. A normal chatbot can explain to you how to fill out a tax return. An agent would try to fill out the forms itself. This shifts the role of AI from an advisor to an employee who acts.

This is precisely why Agent Mode is economically interesting for companies. Tasks that consist of many small, tedious steps could partly run without humans. That is why software companies have been advertising almost every new product with the word “agentic” since around 2024. Whether the systems are really reliable enough is the central open question.

With freedom of action, the risks also grow. An incorrect text suggestion is harmless, a deleted file or an accidentally submitted order is not. For this reason, many providers have the agent ask for confirmation at critical steps. This is called a confirmation step: the human must agree before something final happens.

The loop of thinking, acting, and checking

Technically, Agent Mode is not a new AI model but a loop built around a familiar language model. A language model is a program that has learned to continue texts and can thereby answer questions. In Agent Mode, it is called repeatedly. Each call answers the question: What is the next sensible step?

If the model decides on an action, it outputs it in a fixed format, for example “search for X.” A control program actually carries out this action and sends the result back. The model reads the result and plans the next step. This round of thinking, acting, and evaluating repeats until the goal is reached or an upper limit of steps has been reached.

The tools available to the agent are determined in advance by the developers. Typical ones are web search, file access, a calculator, and a command line, i.e., the text input through which one gives a computer direct commands. A common misconception is that the agent can do anything on the internet. It can only do what it has been explicitly allowed to do. The biggest weakness remains error propagation: an incorrect intermediate step often poisons all subsequent ones.

Agents in programming tools and browsers

Today, Agent Mode is most widespread in programming. Development environments such as GitHub Copilot, Cursor, or Claude Code offer a switch with exactly this name. With it, one describes a change in everyday language, and the tool edits multiple files, runs tests, and corrects its own errors. Programming fits well because results can be checked automatically: the code runs or it doesn’t.

There are also browser agents that click on web pages and fill out forms, as well as assistants that book appointments or sort invoices. In news reports, such systems often appear under the buzzword “agentic AI.” For those who want to distinguish the terms: Agent Mode refers to the specific operating mode in a product, agentic AI refers to the entire subject area.

For users, a sober look is worthwhile. Agents save time on clearly defined tasks with verifiable results. On open-ended tasks without clear success criteria, they often run into dead ends. It is advisable to narrowly limit access and to check important intermediate results yourself.