Autonomous Agents
Autonomous agents are AI programs that don't just answer a goal but work through it in their own steps. They plan, use tools such as search engines or calendars, and check their own intermediate results.
Most AI programs that write text work according to a simple pattern: you ask a question, they answer, then they wait. Autonomous agents go a step further. They receive a goal instead of a question, for example: “Find the three cheapest train connections to Hamburg on Saturday and add the best one to my calendar.” After that, they act in multiple steps on their own, without a human triggering every step. They search the internet, read results, discard unsuitable hits, and if necessary start a new attempt. The word “autonomous” doesn’t mean that no one controls them, but only that they decide independently between the task and the result.
From answer-giver to task-taker
The difference sounds small, but it changes what AI is actually good for. A program that only answers is an advisor. An agent is an employee to whom you hand over a task. This is exactly where many tech companies see the next big market: software that takes over work steps instead of merely explaining them.
Economically, this is interesting because tasks are worth far more than information. A company pays little for good advice on booking a trip. It pays significantly more for a system that handles thousands of bookings without staff. That’s why agents have appeared in almost every major AI company’s product announcement since 2024.
At the same time, this creates a new risk. Whoever is allowed to do something themselves can also cause harm themselves. An agent that misunderstands an order has already triggered it before anyone can object. Errors therefore no longer stay on the screen, but have effects in the real world.
The loop of planning, acting, and checking
At the core of an agent is a language model, that is, an AI system that has learned from vast amounts of text to formulate sensible sentences and plans. Around this model, a loop is built. The model first considers which step makes sense next. Then it carries out this step. Afterward, it reads the result and decides whether the goal has been reached or another step follows.
Acting here means using a tool. Tools are small programs that the agent is allowed to call: a web search, a calculator, a calendar, a database. The agent doesn’t write flowing text to the user, but a structured call that the software can execute. The result comes back into the loop. In this way, many small text-based decisions turn into a chain of actions.
For this, an agent needs a memory, because after twenty steps it still has to know what it found out in step three. It usually stores intermediate results in a notes file or database and reads them back in as needed. A common misconception is that agents are a completely new kind of AI. It’s the same technology as with a chatbot, just with tools, memory, and a repetition loop built around it.
Agents in products and headlines
Today, agents are most visible in programming. Tools like Claude Code or GitHub Copilot Agent receive a task description, independently change multiple files, run tests, and fix bugs they find. Similar systems control a browser, fill out forms, and compare prices. In companies, agents take over sorting support requests and preparing refunds.
In the news, you usually encounter the term in two contexts. First, in revenue hopes: providers expect customers to pay more for completed tasks than for chat access. Second, in reliability problems, because in tests agents still frequently fail at long chains of tasks. If each of twenty steps only succeeds 95 percent of the time, the overall task is often already lost.
In practice, this means: trust agents where mistakes are easy to undo. For payments, contracts, or emails to strangers, human approval makes sense. This intermediate step is exactly what the industry calls “human in the loop.”