Lots of what software can do is reachable only by clicking and typing in its window, so automating real work means an agent, a model that picks actions, must read the screen, act, and keep a long workflow on track.
NorthCUA builds varied practice tasks automatically, runs them in a harness that records every decision and checks whether the task actually got done, then trains on the verified runs by imitation, followed by reinforcement learning, which is trial and error scored by success.
Its largest model, 397 billion parameters (its size measure), leads open-source rivals on screen, desktop and web tests, and training drew on over four million recorded interaction steps. The loop between doing and retraining is the reusable part.