Nvidia’s idea of a real agent is AI that keeps working on its own for as long as 12 hours after being given a task. It is different from AI that gives an answer when a person asks a question and then waits for the person to revise the question and ask again.

Kim Sun-wook (김선욱), a managing director at Nvidia Korea, said at the Digital Insight 2026 conference in Seoul on Sept. 29 that “an agent runs for 12 hours, 5 hours once you give it a task.” The event was held by DigitalToday under the theme “Agentic AI.”

Kim added that “you have to run all the connected things, like tools, together,” and that “what runs in orchestration is the real agent.”

Companies are already looking for resources to run such agents. “We get a lot of inquiries asking if they can rent computers because they don’t have computing resources,” Kim said. “That shows how much companies are using agent AI.”

Kim described an agent as a structure in which a harness is attached to an LLM. If the LLM is the brain that understands language, the harness is the reins that direct that brain. The harness remembers user records to refine questions and connects files, computers and tools to act in place of a person. For the brain role, users select from options such as models made by Nvidia and Google to fit their computing environment.

The harness runs on CPUs, he said. “LLMs run on GPUs, and the harness runs mostly on CPUs,” Kim said. “That is why CPUs have become quite important.” He said changes in how agents are trained have also increased the CPU’s role.

“In the past, we trained the LLM itself, but now we train the agent,” Kim said. CPUs run the process of repeating tens of thousands of times while receiving feedback on whether the agent did its job well. He said that is why Nvidia built separate CPU racks. Nvidia currently makes and uses Vera, an Arm-based CPU.

Tasks such as storage security and data processing that CPUs cannot handle are taken over by DPUs, data processing chips called Data Processing Units. Kim said DPU refers to a chip Nvidia made to separate out data processing that CPUs used to handle. It parses and processes data through a high-performance network interface and efficiently transfers data to GPUs and CPUs.

Next, the foundation that lets the agent keep working is memory. The agent works by loading its assigned mission and its history into memory. “What people are good at is because they have memories accumulated since childhood,” Kim said. “That is why memory is also very important.”

Even with these chips, an agent is not completed, he said. Kim said that is why some companies only make models while others only make harnesses. Nvidia therefore provides an agent toolkit that includes its in-house LLM Nemotron, data acceleration libraries and safety devices so those components can be assembled. “You can use this or use other models,” Kim said. “If you change only what you are good at, you get another result.”

Ultimately, Nvidia sees agents as an infrastructure issue. Kim said agents can run only when CPUs, GPUs and storage servers are tied together into a single infrastructure. He said electricity drives that infrastructure. “You can run agents within the range that energy allows,” Kim said. “How much money it is comes down to how much electricity you put in and how many tokens you extracted.”

The next stage Nvidia is looking at is robots. Kim said robots are like moving people, so the burden on agents is far greater, adding, “We have to build an even more complex system than this.”