What structure should the next generation of AI-native applications have?
Until now, most software has been designed on the assumption that people operate computers through a graphical user interface. Once AI agents begin operating applications autonomously, however, people and AI no longer need to share the same interface. Structured controls such as a CLI or API are the easiest for AI to use. For people, what matters is being able to understand intuitively what changed as a result of the AI's actions and to verify those changes.
Andrej Karpathy, who worked at OpenAI and Tesla and now conducts research and development at Anthropic, has likewise argued that collaboration with AI will increasingly be based on a division of roles in which AI generates and people verify, making the speed of that verification loop crucial. He has also described visual information such as a GUI as “a highway to your brain,” useful for auditing AI output.
Architecture already has a rich set of representations whose meaning people can grasp visually: plans, three-dimensional models, meshes, temperature distributions, and airflow. We believe these will remain the best interfaces for intuitive human verification in an AI-native era.
Montelo embodies this idea. Whenever AI operates a model or analysis condition through a CLI behind the scenes, the GUI updates in real time. When the AI places air-conditioning equipment, the change immediately appears in the 3D model. When it generates or refines a mesh, the change becomes visible on screen as it happens. Although a CLI drives the process underneath, the state changes before the user's eyes as if someone were operating the application in front of them.
The reverse is also true. When a person moves equipment or changes a setting in the GUI, the modification is immediately reflected in the underlying context referenced by the AI.
AI acts on the system through the CLI. People observe it through the GUI and provide corrections and checks. They use different interfaces, but they always refer to the same shared state.
Separating the operating surface for AI from the observation and correction surface for people, while allowing both to share the same underlying state, is the fundamental structure of an AI-native application as we see it.