Nova Act is a "research preview" because it still makes dumb mistakes. Our lab is working on that: labs.amazon.science/careers

For an agent to add value it needs to work in the background, reliably, with zero prompt engineering. Computer use models aren't good enough today. Wrapped by an application, prompted by a developer, they have much more potential.

High-level prompts with long trajectories show up *all the time* in demos and multimodal agent evals. They imagine a user typing into an empty textbox - like Adept's original "find me a house" demo (Aug 2022!). But it's hard to save time by watching your computer work.

Two risks we've taken with Nova Act: focusing on short trajectories, and launching an SDK as the first interface.