Agentic Tasks
Agentic tasks are assignments that a computer program does not complete with a single answer, but rather through many self-chosen intermediate steps. The program plans, uses tools such as search engines or files, and checks along the way whether it is getting closer to the goal.
You can give a chat program two very different kinds of assignments. The first type is a question that fits a single answer: “What is the capital of Norway?” The second type is a goal without a fixed path to get there: “Find the three cheapest train connections to Oslo and enter them into a table.” Tasks of the second kind are called agentic. The program has to break the assignment down into steps itself, gather information along the way, and decide when it is done. The word comes from the Latin “agere,” meaning to act: the program does not just answer, it does something.
From answer-giver to task-taker
Until now, AI programs have mainly been judged on how good their individual answers are. With agentic tasks, something else counts: whether a finished result is there at the end. Economically, that is a big difference. A good answer saves a person a few minutes of research. A completed task can take an entire hour of work off their hands.
This is exactly what many technology companies are currently pinning their expectations on. That is why firms like OpenAI, Anthropic, or Google constantly talk about agents in their announcements. Anyone selling software that completes tasks instead of just delivering text can charge significantly more money for it. Investors and analysts are following this development closely, because it plays a role in determining the industry’s future revenue.
But there is also a hard catch: errors add up. If a program is correct 95 percent of the time at each individual step, a chain of twenty steps is only completely error-free in about one out of three cases. This is exactly where many agentic systems still fail today. Reliability over long chains is currently a bigger problem than intelligence at the individual step.
The loop of planning, acting, and checking
Technically, almost every agentic task is built on a loop. The language model, meaning the program that understands and generates text, first thinks about a next step. Then it carries out that step by calling a tool. A tool can be a web search, a calculator, a calendar, or a program that opens and saves files.
The tool’s result goes back to the model. So it reads what its own action accomplished. Then it decides: continue, correct, or stop. This round repeats until the goal is reached or a limit kicks in. You can think of it like an intern who briefly checks their notes after every step of work.
Two components make the difference here. First, tool access: without it, the model stays trapped in its own text window and cannot change anything in the world. Second, memory of the process so far, because by step fifteen it still needs to be clear what came out of step three. This should not be confused with pure automation: a classic script follows a predetermined sequence, whereas an agent determines the sequence itself.
Agents in the browser, in the editor, and in the quarterly figures
Agentic systems are furthest along in programming. Tools like GitHub Copilot, Claude Code, or Cursor receive a bug description and then work through a project independently. They read files, change code, run tests, and fix what the tests flag as wrong. This works particularly well here because the test automatically tells whether the result is correct.
There are also agents that operate a browser and fill out forms or compare prices for you. In companies, they take on tasks such as sorting support requests or compiling reports from multiple data sources. In news reports, English terms like “agentic AI” or “AI agents” often show up in this context.
If you read such reports, it is worth asking a critical question: how often does the system really accomplish the task on its own? Providers cite success rates for this from tests like SWE-bench, which contains real programming problems from open-source projects. These figures were long in the low single-digit percentages and have since risen considerably. Still, in practice it usually holds true: a human checks over it at the end.