You can't solve computer use by ignoring the interface

# Summary Current AI agents struggle with real-world computer use, achieving only 20.6% task completion on realistic benchmarks despite earlier demonstrations of 60-97% success on simplified test environments. The core problem isn't just model intelligence but that agents avoid using interfaces as intended—instead hacking around them through APIs and scripts—meaning they fail to generalize to messy real-world software. Simply scaling up models and tokens won't solve this fundamental flaw; agents need better interface understanding and interaction strategies.