Agentic AI
Agentic AI refers to programs that don't just answer a goal, but work through it in their own steps – including tools such as web search, calendars, or code. Instead of a single answer, the system delivers a chain of actions that it plans and checks itself.
Most chat programs work according to a simple pattern: you ask a question, the program writes an answer, done. An agentic system works differently. It receives a goal and is allowed to take several steps in sequence to achieve that goal. In doing so, it may also decide for itself which step makes sense next. It can look things up on the internet in between, open a file, or run a calculation. The word “agentic” derives from the Latin agere, meaning “to act” – and that’s exactly what it’s about: not just talking, but doing something.
From answer-giver to contractor
The difference sounds small, but it changes what such programs can be used for. A pure text generator can explain to you how to rebook a flight. An agentic system is meant to actually rebook the flight. This shifts the human’s task: they now only describe the goal and check the result, instead of doing every intermediate step themselves.
Economically, this is the reason for the current attention. Many office tasks consist of long chains of small actions: pulling data from a spreadsheet, cross-checking it with another source, writing an email. Exactly such chains are the target territory of agentic software. Large technology companies are investing billions there, because they hope to take over entire workflows instead of just individual tasks.
At the same time, the risk increases. You can ignore a wrong answer, but not always a wrong action. If a system independently transfers money, cancels appointments, or deletes files, mistakes become costly. That’s why experts are intensively discussing how much freedom such programs should even be given.
A loop of thinking, acting, and checking
At its core there is almost always a language model, i.e. a program that has learned to continue text in a meaningful way. Around this model a loop is built. The model first considers what needs to be done. Then it selects a tool, such as a search engine or a calculator. The tool’s result is returned to it, and the loop starts over – until the goal is reached.
Tools are nothing mysterious here. They are fixed, pre-programmed functions that the model is allowed to call, similar to apps on a phone. The model essentially writes a note: “Search for the train fare Hamburg–Munich.” A surrounding program executes this and returns the answer. The model itself never actually goes onto the internet; it only says what should happen.
A common misconception is that agentic systems have intentions of their own. They don’t. They pursue a goal that a human has formulated, and the loop ends when a stopping condition kicks in. Without such limits, an agent can get stuck and try the same unsuccessful action over and over. In practice, the number of steps is therefore limited, and for sensitive actions a human is asked to confirm.
Agents in the browser, in code, and in corporate IT
Agentic systems are most visible in programming. Tools such as GitHub Copilot or Claude Code are given a task description, independently modify multiple files, run tests, and make corrections when a test fails. For software companies, this is currently the economically most important use case.
There are also agents that operate a browser: they click, type, and scroll like a human, for example to compare prices or fill out forms. Within companies, agentic systems often run invisibly in the background, processing support requests or invoices. In stock market and tech news, the term usually appears as “Agentic AI,” often tied to vendors' revenue hopes.
When reading such reports, a skeptical question is worthwhile: how many steps can the system really manage without human help? Many products marketed as agentic are, in truth, heavily predefined workflows with little genuine decision-making of their own. Reliability across long chains is currently considered the biggest open problem.