According to WatchTower AI monitoring, Huang Renxun elaborated on the architecture of the "AI Agent" during his speech at GTC Taipei 2026: a large language model is responsible for reasoning, inference, and planning, while an external orchestration engine (harness) acts like an operating system, connecting the model to tools such as spreadsheets, browsers, and databases, and managing working memory and long-term memory. He defined this as a new computing paradigm—where the past was "launch the app, click input," the future is "describe intent to AI, AI generates code, calls tools, and produces results."
Huang Renxun demonstrated multiple Agent instances live: generating full application code using natural language prompts; inputting a description to generate a dynamic particle animation of Taipei 101→GTC Taipei 2026→NVIDIA logo; and even taking a photo of a lost remote control battery clip, where the Agent directly called a CAD tool to generate a 3D-printable replacement part file. He mentioned, "Here we are using Claude Code, but Codex performs equally well."
Addressing the claim that "Agents will cause software companies to go bankrupt," Huang Renxun rebutted once again: "On the contrary." He believed that in a world no longer constrained by human capacity, a massive number of Agents would use more software tools than humans, making it "the best time to run a software company" — but the software must be presented in a way that Agents can invoke. He also announced that NVIDIA's CUDA X library is now fully open to Agents, with Agents using these libraries even more efficiently than human developers.
