You can't solve computer use by ignoring the interface
Summary
The Steelman Labs post argues that solving computer use is not achieved by simply scaling models or token budgets, highlighting that interface design and interaction are critical bottlenecks. It presents OSWorld-V2 findings showing low long-horizon task completion and contends that agents often bypass the user interface, leading to suboptimal and brittle outcomes. The piece proposes a steelman approach that separates planning from execution with a vision-first, UI-aware agent capable of interacting with any interface a human can use.