Agentic AI Workloads

When you ask an AI chat program a question, what usually happens is this: the program computes once, writes an answer, and is done. With agentic tasks, it works differently. Here the program is given a goal instead of a question, for example: “Find the three cheapest flights and book the right one.” The program breaks the goal down into steps itself, searches the internet, compares options, corrects itself, and keeps working. It thus acts over many rounds without a human triggering every single step. The word “workload” here refers to the amount of work this generates on the computers of a data center.

Why a single question suddenly turns into a thousand computing steps

A normal chat response costs the provider a fraction of a cent. An agentic task, by contrast, can trigger hundreds of individual model calls. The program thinks, calls a tool, reads the result, and thinks again. Each of these steps is its own computing task for the graphics cards in the data center. Resource consumption therefore doesn’t rise by ten percent, but often by a factor of a hundred.

This is exactly what makes the term interesting to investors and market observers. Chipmakers and data center operators now explicitly justify their growth forecasts with agentic applications. The reasoning behind this is simple: if every user request triggers far more computing steps, the world needs far more hardware. Whether this demand will actually materialize at the forecasted pace is one of the industry’s big open questions.

At the same time, providers' pricing models are changing. Many no longer sell their AI per response, but as a subscription with usage limits or per completed task. The reason is that the cost of a single request can hardly be predicted anymore.

A loop of thinking, tools, and memory

Technically, such a workload consists of three building blocks. The first is the language model, i.e., the trained AI system that processes and generates text. The second is tools: search engines, databases, a calculator, a code interpreter, or interfaces to other programs. The third is a memory in which what has happened so far and what is still open are recorded.

These three run in a loop. The model decides which tool it needs next. The tool’s result is fed back to it and moves into memory. Then it decides again. The loop ends when the goal is reached or a predefined limit on the number of steps is exceeded. Such limits are necessary because otherwise an agent can end up going around in circles.

The comparison with an intern holds up fairly well here. You give it a goal, it makes calls and does research on its own, and in the end reports back with a result. But just as with the intern: a mistake in the third step poisons everything that follows. That’s why check steps are built in, for example a second AI that reviews intermediate results. A common misconception is that agentic systems are especially “intelligent.” The underlying model is the same one; it’s simply allowed to make multiple attempts and use tools.

From coding assistants to quarterly earnings

Today, agentic workloads are most widespread in programming. Tools like GitHub Copilot in its newer versions, Claude Code, or Cursor edit multiple files on command, run tests, and fix whatever isn’t working. The “deep research” features of major chatbots also belong here: they search web pages for minutes at a time and then deliver a report complete with sources.

In companies, they show up in customer service and accounting. An agent reads an incoming invoice, checks it against the purchase order, and submits it for approval. In everyday life, you encounter the principle in assistants that independently coordinate appointments or complete orders.

In the news, the term mostly comes up in technology companies' quarterly reports. When Nvidia, Microsoft, or Amazon talk about agentic workloads, they are justifying rising data center spending. The term should be distinguished from “automation” in the classic sense: there, a human predefines the sequence of steps; here, the system decides it on its own.