Computer-use agents have surged from 42% benchmark performance in early 2025 to 85% today — clearing the 72% human baseline on OSWorld. But the real story isn't the score: the model is no longer the bottleneck, and the teams winning in production are the ones who figured out verification, escalation