Agentic Transition

For a long time, most AI programs worked according to a simple pattern: you ask a question, the program answers, and then it’s done. For a few years now, this has been shifting. Newer systems are given a goal instead of a question and work through it themselves in multiple steps. They break down the task, call on other programs, check their intermediate results, and continue. This shift from a responding system to an acting system is called the agentic transition. The word comes from the Latin “agere,” meaning to act.

From tool to coworker

The difference sounds small, but it changes the usefulness enormously. A responding system is like a reference work: it delivers knowledge, and the human does the rest. An acting system also takes over the work steps in between. Instead of explaining how to book a trip, it searches for flights, compares prices, and fills out the form.

For companies, this is the actual economic point. The value of an AI no longer rises only with the quality of its texts, but with the length of the task chain it can complete on its own. Companies are therefore increasingly selling not answers, but completed processes. This explains why, since 2024, quarterly reports and tech conferences have constantly been talking about agents.

At the same time, the risks grow along with it. A wrong sentence is annoying, but a wrong order costs money. Anyone who lets a system act must draw boundaries and catch errors early. That is why regulators are discussing above all what a company is liable for when the software has made its own decision.

Planning, calling tools, refining

At its core, there is still a language model, that is, a program that has learned to generate sensible continuations of text. What’s new is the loop built around it. The model first writes a plan in text form. Then it carries out the first step, looks at the result, and decides on the next one. This round repeats until the goal is reached or a limit is exceeded.

A model can only act through tools. Tools are connected programs, such as a web search, a calculator, a database, or an email dispatch. The model doesn’t write flowing text, but a structured call that another program executes. Its return value ends up back in the model’s text. This is how language turns into actual effect in the world.

On top of that comes a memory that retains intermediate results across many steps. Without it, the system would forget after ten steps what it originally wanted. A common misconception is that an agent is a completely new kind of AI. In fact, it is usually the same model as in the chatbot, supplemented with a planning loop, tools, and memory.

Agents in the browser, in code, and in support

The development is most visible in programming. Tools like Claude Code or GitHub Copilot are given a task description, independently change several files, run tests, and fix what fails. Something similar exists for the browser: the system clicks its way through web pages and fills out forms. In customer service too, systems don’t just answer questions but directly trigger refunds.

In news articles, you can recognize the topic by terms like “agentic AI,” “AI Agents,” or “Computer Use.” Mostly, it’s about the question of how reliably such systems can accomplish longer tasks. This is often measured by how long an agent can work before a human has to step in. This time span has grown considerably in recent years, but rarely extends beyond a few hours.

You often encounter the agentic transition in everyday life without noticing. When an app does its own research, suggests appointments, and writes confirmation emails, exactly this principle is at work behind it. The interesting question is less whether this is technically possible. It is rather how much control one gives up, and at which point a human still needs to confirm.