At GTC Taipei 2026, Jensen Huang laid out a systematic architecture for what he calls "AI agents": a large language model handles thinking, reasoning, and planning, while an external orchestration engine — a "harness" — acts like an operating system, connecting the model to tools like spreadsheets, browsers, and databases, and managing both working and long-term memory. He framed this as an entirely new computing model: instead of launching an app and clicking through inputs, the future is describing your intent to an AI, which then generates code, calls tools, and delivers results.
Huang demonstrated several live agent examples: generating a full application from a natural language prompt; creating a dynamic particle animation of Taipei 101 transforming into the GTC Taipei 2026 logo and then the Nvidia logo from just a text description; and even snapping a photo of a missing battery cover from a remote control — the agent then called a CAD tool to generate a 3D-printable replacement part. "We're using Claude Code here," he noted, "but Codex performs just as well."
As for the fear that agents will put software companies out of business, Huang pushed back hard: "Quite the opposite." He argued that when the world is no longer limited by the number of humans, a vast army of agents will use far more software tools than people ever could. "This is the best time to be a software company" — but only if that software is built in a way agents can call. He also announced that Nvidia's CUDA X libraries are now fully open to agents, and that agents use these libraries even more efficiently than human developers.
The "harness as OS" framing is interesting, but IMO it sidesteps the harder organizational question: who builds and governs that harness at enterprise scale, across every function? Huang describes the mechanism, but he doesn't describe the practice.
An army of agents calling more software tools than humans ever could is only a good outcome if there's something coherent underneath them. Without that, you get faster chaos instead of the faster outcomes you're looking for.